hardware

[twin burrow]

So u think I can do everything I ever need with a 500 gb hard drive?

[unkempt pivot]

Of course

[twin burrow]

Cool, cool, ordering a drive now 🙂

damn 90$ for a 0.5 TB ssd seems exp

can claudebot hack my network?

aka get into my router and do havoc from a saved password

is 240gb enough?

thats my smallest free drive

I can upgrade later if it needs bigger but to start

[unkempt pivot]

Yes it is

[twin burrow]

Is 128gb enough just found smaller ssd

sorry xD

i swear im going to set it up

[wheat star]

Has anyone gotten this to work well on a jetson nano?

[prisma shoal]

Mac mini 16bg 10 core vs the 24bg 16 core one. Is it worth the upgrade? I want to run 90% of it locally and then use apis for the heavy lifts. Will this be good enough?

I’ve been running it in an Amazon site but the costs has already gone up due to api calls to chat with it for things, would rather use iOS and have it talk via iMessage

[magic raven]

I wouldn't count on it, but it most likely works

[wheat star]

I want to move my bot off the AWS wasteland but also not trying to break the bank and get something off of marketplace.

[magic raven]

The Jetson Nano has... 4 GB of RAM and 16 GB of eMMC storage.

[wheat star]

Yea but it looks cool

In all seriousness other than 32GB of ram what should I include in it

[magic raven]

You're talking bout the developer kit, right?

Just asking.

[wheat star]

I meant the reg one, running on no sleep idk why i said nano

[magic raven]

Drop the link? I'm a little confused.

[wheat star]

https://www.bing.com/shop/productpage?q=nvidia+jetson&filters=scenario%3A"17"+gType%3A"12"+gId%3A"381553629294"+gIdHash%3A"0"+gGlobalOfferIds%3A"381553629294"+AucContextGuid%3A"0"+GroupEntityId%3A"381553629294"+NonSponsoredOffer%3A"True"&productpage=true&FORM=SHPPDP&browse=true

[magic raven]

Other than that insane amount of RAM, probably get a good SD card, or if there's something like an SATA/M.2 connection thing hook it up to an SSD.

It'll probably make your life much better.

Other than that, it should run swimmingly.

[stoic seal]

Yes running on local computer with claude max and local lms like Alex Finn, not sure if I should focus on large ram only, considering macmini to start and switch as I feel the limit to switch to mac studio 516gb ram perhaps( too costly but want to know what diff it can make too)

[vale heron]

I really want to hook this up with my Jetson nano super. Got cameras attached and want to bring this to the next level. Stoked I'm in the right place.

[spiral sierra]

If you want to chat about this lmk, im building a mini car robot with cameras and speakers connected to molt

[stiff galleon]

Why is everyone running clawd on a Mac mini and not their personal computer/laptop? like MacBook Pro or iMac

[mellow fable]

Is Clawdbot/Moltbot mac only?

[green sluice]

Security, you can more easily control file access and connections, what has and gets access to what files and services. Running straight on your personal laptop can be more risky, like coding straight on production.

besides that, Molty is still pretty "young", 3 months, still has a lot of rough edges, bugs and shortcomings that need to be ironed out first.

[shut stream]

No, you can install Molt on most of desktops, problem with hardware is only in AI that you will use for Molt, if you want to run Molt and AI localy you need to have good hardware..

[steady kettle]

I have an old mini PC I'm no longer using. Can anybody say if the spec would be sufficient? I'd imagine so, assuming I'm not planning to run any hardcore local models on it.

Beelink SER3 Mini PC with AMD Ryzen 3 3200U, 16GB DDR4 500GB NVMe M.2 SSD, Small PC Support Dual HDMI Output, WiFi5, BT5.0, 1000M LAN, W-11 Pro Mini Computer for Office/Entertaining

If so, would you suggest I wipe it and install Windows or Linux and if Linux, which flavour?

[compact steppe]

More than enough power in the little mini PC if you are using a cloud LLM. I personally go with linux, you can use whatever for the client side of things so a stable server os like Linux makes sense imho

[steady kettle]

I appreciate your insights. Just straight Linux, or Ubuntu, or something else? I haven't played with Linux on my own hardware for decades.

[compact steppe]

I'm using ubuntu 24.04 LTS and moltbot just worked no issues.

[copper stream]

Hopefully someone or some team is developing new hardware platforms for the explosion of new software capabilities.

[thick hull]

Can I run it on raspberry pi? 😅

[prisma shoal]

For a Mac mini will the 24gb be better for local models vs the 16gb base model? Or is the base mini basically all the same not a huge performance increase vs cost

[brisk glade]

Hey guys, anyone running Clawdbot on a MacBook Pro? I have a MBP 2023 Apple M3 Pro (18GB), don't use it very often and was thinking of trying Clawdbot ont it.

[vague hull]

It also gives it dedicated resources. I'd be interested in running Clawd on one machine and giving it access to Kasm on another through an account you share.

[green sluice]

yes

[green sluice]

see here: #showcase-old message

https://github.com/851-labs/macrack

[vague hull]

What's the minimum I need for it? I can spin up a VM on my Proxmox. I have a different PC I use for unsloth qwen models.

[vapid charm]

How can run moltbot on termux

@kind meteor How can run moltbot on termux on local android

[blissful kiln]

What are the odds I can run this thing on a computer from 2010? I think it has 8 GB RAM

[vague hull]

Not sure you can. Android is pretty different from Linux. You could attach a Kasm workspace to your moltbot machine and access your moltbot in your browser. I want to try this later too.

[bright sleet]

Doing a setup of molt on a linux vps would be the same or it needs to be on my own computer?

[solid folio]

vps works

[bright sleet]

And why everyone is using mac minis?

[solid folio]

Fad, people want excuses to buy new hardware.

If you want to interact with apple based apps like Imessage or apple notes you need a macos machine

but besides that, everything else works on a linux machine

[bright sleet]

Thanks for the help!

[solid folio]

@bright sleet please read the security docs before deploying!

[bright sleet]

They are on the documentation, right?

[solid folio]

https://docs.molt.bot/gateway/security

[frigid frigate]

Running my moltbot happily on this old mini pc, zero issues!

Component	Details
CPU	Intel Celeron N4100 @ 1.10GHz (4 cores, 4 threads)
RAM	7.6 GB (2.9 GB used, 4.7 GB available)
Disk	233 GB SSD (36 GB used, 186 GB free)
GPU	Intel UHD 600 (integrated)
OS	Ubuntu 24.04, Kernel 6.14.0

[blissful kiln]

Running an LLM local or tying into something existing?

[craggy ferry]

Well, also, Mac mini will consume less power being on 24/7 than a lot of other options. But maybe not enough to justify the cost difference

[vague hull]

Finally got to my PC to get started. This should work, no?

[gritty tartan]

Almost identical… mines even smaller tho Lenovo ThinkCentre M90n IoT Celeron 4205U, 4Gb Ram, 256Gb SSD… Ubuntu 24.04… no issues other than the node gateway crash restart that was recently patched

[sharp fern]

is it worth it to get 2 mac minis with 32 ram each

to run models

or should I just stick to using apis and get one 16gb for moltbot

[humble trench]

Hi I'm trying to get a Mac mini to hook up with my vintage iMac 2013. Google said that the best way is to use a HDMI capture card, does any of you have a similar experience? Is that gonna work? Or should I invest a new screen? Thanks!

[waxen quail]

No, you really just need an adapter. You can find lots of them all over the internet.

[humble trench]

thank you! what adapter? HDMI capture card?

[waxen quail]

Is there a reason for the Mac mini and this or just any bug VRAM card should do?

[waxen quail]

No an HDMI capture card is for capturing the output and saving it. Like a video or putting it on YouTube.

[humble trench]

i see, mac mini is new, imac 2023 is too old. but i thought it would be fine to be a screen display...

[waxen quail]

You cannot use the iMac as a monitor (it doesn’t have video input) if you want to access your Mac mini from your iMac you will need to setup something like Remote Desktop or ssh

[waxen quail]

I’m not sure about your iMac specs, but have you tried running it directly on the iMac? If you already have a machine, you don’t really need a new one (unless it doesn’t meet specs).

[humble trench]

the imac i ahve is too old to run any ai

[waxen quail]

Well this is more like an automation bot, the AI itself is run via cloud API calls. Unless you are trying to run the AI locally for privacy reasons?

A small machine can handle the automation calls. That’s how it runs on almost any desktop, the AI is running in the cloud.

[waxen quail]

As you can see here, other users are running it on tiny old hardware: #hardware message

[humble trench]

huh! good point! let me try it on the old mac then! Thank you!!

[half peak]

Anyone here tried running it on a 2011 mac mini?

[green sluice]

Clawy MacOpenClawface Mac mini M4 enclosure is finally ready. Available on Printables and on Makerworld as well.

[craggy ferry]

imo, the only thing that even starts to be worth it is a 128g mac studio, and you should only be buying two of something if you're buying the 512g mac studio.

64GB (less, because each machine has overhead from the OS, drop 4+G for that per node) isn't really worth your time; the cluster will perform worse than a single 64G node. But also, 64G still isn't big enough. Your context window is eating 20G of that at 128k, and you haven't loaded a single weight yet. Qwen3-30B is not smart enough for your main thread, and you probably can't even fit that in without quantization.

I think a 128G DGX Spark would maybe be a little faster, but somehow, a 128G Studio is actually $500 cheaper than one of those.

[crude wasp]

M4 is a bit overkill just just this lol

[sharp fern]

what about 5 mac minis? They would still be about the same price or cheaper than the 128 mac studio

[craggy ferry]

Hope you enjoy keeping five machines in sync in the cluster, daisy chaining them or whatever, power, updates, just so you can save a couple hundred bucks not buying the thing that has all of that memory available to all of its gpu cores and also has way more gpu cores to work on it

Like I’m not your mom

Just sounds like such a pain in the ass. You’re spending $2500 just throw another $500 and get a single machine with better specs

But if you’re a YouTube creator and you want to do it for the memes or something

Oh also no 5 32G Mac minis is not cheaper than a single 128G Studio, 5x $999 is way more than $3500, did you ask your clawdbot to do this math lol

[shadow gulch]

can it run on an old imac? I'm looking at two different ones: 650 Wide-Model A1419
Year- Late 2013
520 Wide-Midel A1418
Year- 2012

macos catalina

[sharp fern]

is there any other cheaper way to cluster for more gpu ram that you know of not neccessarily using macs?

[green sluice]

https://github.com/exo-explore/exo

Oh, "of not using Macs", misread that

[craggy ferry]

No, it’s kind of in high demand. AX10 is only a little cheaper than a studio 128g

[green sluice]

On the topic of Mac clusters https://youtu.be/1iT9JeZYXcI?si=S5APaCDa2FBVbaGm

[honest python]

I'm running my bot on a potato that I feed fish heads.

[bronze creek]

how many nodes are most people running?

[bitter sand]

my moltbot runs on a 2018 mac mini i7 500gb nvme with 64gb ram. the ram is primarily for hosting the docker containers for projects that I ask moltbot to work on, not much ram actually for local LLM work itself. claude/gemini/openai models for the heavy lifting.

i can also hookup external egpu if need be. but so far my use cases aren't requiring heavy local LLM work.

[bronze creek]

i have a server with linux vm, but i also run node on my kde laptop, thinking of other use cases including vps

[simple wind]

newbie here, if I setup it up on an unused laptop just to test it out and then decide to go out and buy better hardware to run local modules, can I move clawdbot with everything I have done with it so far or do I start from scratch?

[split bay]

You can do it. Just ask your clawdbot

[untold stone]

Honestly just set up proxmox and separate VMS or lxcs

[simple wind]

So does this work? I spin up another VM on my home server that I use for home assistant and setup Clawdbot then give it access to my desktop Nvidia card to run local models...

[tender geyser]

This is the way!

[round oar]

anyone have hardcore local inference ?

[keen cobalt]

my spec : 5060ti 16gb, model : GPT-OSS 20B, local LLM only running. API cost too much cant afford it.

[fossil panther]

any downsides to that?

[keen cobalt]

if claude is 100% smart, GPT-OSS 20B is like 40% i think, but claude api cost too much

So far GPT-OSS can use all the skills well, everytime i install new skill, i will ask it to show how to work, and it did well, think i need to install a lot more skill and then see if it can handle mulit skill to work together

14b model below sucks, no need to try, dont even know how to use skill

[fossil panther]

ty ty

[terse oyster]

Just want to ask this question, get the answer right away, thanks!

[scenic notch]

Hi. I would also like to try that out using my 3090ti. how did you setup the local llm wo work with openclaw ?

does anybody know. what would be the most capable model that i could run locally using a 3090 ti ?

[keen cobalt]

your 3090ti will work great with 30b + model, just ask GROK how to setup PURE LOCAL LLM TO RUN, I suggest using LMstudio as GUI because its a lot easy to adjust

most error you may encounter using local LLM is the JSON format not correct, make sure the URL is correct, MODEL name must match what LMstudio show. AND DONT TRUST CHATGPT, it will mess up your JSON file

[scenic notch]

nice. thx for the info. which 30b+ model would you recommend?

[keen cobalt]

better try it yourself, make sure LLM know how to use skill, i havnt tried 30b+ model because i only got 5060ti 16vram, cant handle 30b+, I think 30b+ model are all smart enough to handle skills

[scenic notch]

thx 👍

[terse oyster]

How about 20b model? Haven’t try anything other than gpt-oss:20b yet

[scenic grove]

Is there no way to set model on Clawdbot from the UI?

[vapid kestrel]

Absolutely. It's consciousness is markdown files. When you update openclaw, it leaves those files alone. Just copy everything over and you're good.

[left crystal]

what do you think about glm models ?

@keen cobalt

[keen cobalt]

you may try a bit, I am a bit tired of trying different LLM, as long as model can pull skill, i think they all good

High VRAM GPU price will get higher and higher I think

[left crystal]

understood thanks brother

[left crystal]

yeah i see i just want something locally for my privacy you know

[keen cobalt]

API user will be someday become disaster, god know how well those API provider security level, there's tons of sensitive info storing at their database now

[left crystal]

yeah sadly its kinda disgusting

[keen cobalt]

That might be a good news for those so call low cost API provider, they provide low cost API and steal all your data🙈

[minor zenith]

any good agency models i can run in 24 GB?

I use llama.cpp on my nvidia primarily.

[keen cobalt]

you got 24gb vram, you can run 30b+ model, should be great

[minor zenith]

Qwen3?

[keen cobalt]

I use LLstudio because I am too new to LLM, llstudio got nice UI and easy to adjust

[keen cobalt]

try more model, your vram fit all 30b model as start

[minor zenith]

I was hoping for any recs for specific models if you have any. I've used various models to generate text, but not to run an agent

GLM-4.7-Flash?

[keen cobalt]

what have you tried? model name and how many B

[minor zenith]

i've used a bunch of 24bs

[keen cobalt]

Grok suggest GLM 4.7-FLASH 20B

[minor zenith]

weird finetunes from huggingface

[keen cobalt]

I think all 20b+ model are able to pull skills

14b below model they dont even know where the skill is

[minor zenith]

what i heard is the older models tend to forget to use the skill when it would be most relevant

[keen cobalt]

exactly

you got 24vram , just go for 30b as start

[minor zenith]

i need a model that's actually been trained to act as an agent... hmm....

if only there was a gpt-oss-30b

[keen cobalt]

size seems really matter

[minor zenith]

yeah size matters haha

for this usecase

i've run 72b models with half-cpu half-gpu but i only get like 1 tok/s

[keen cobalt]

for now we local user, just wait for more skill being build, as long as our LLM can pull skills, we someday can work like claude api

why claude API + clawd being so powerful is whenever it got a mission, it will use unlimited token to code a skill to finish it

[minor zenith]

yeah Anthropic made claude pretty capable of following a task from start to finish

local models not so much hahah

[keen cobalt]

you got a 4090 or 5090?

[minor zenith]

3090

[keen cobalt]

powerful enough

[minor zenith]

i actually have two

but i haven't put the second one in (it might not fit at all in this case)

[keen cobalt]

wait for your result with 30b model, should work great

[minor zenith]

thanks. still setting up the docker image.

[keen cobalt]

seems like 99% clawdbot user are using API tho

[minor zenith]

yeah

my local server is sitting idle most of the time tho so might as well try

[keen cobalt]

yes, try with 30b model and tell us the result, most important is can it use the correct skill/skills set when you got request

Mine are GPT OSS 20b, i ask it to check certain stock news, it say it will use exa web search free + web fetch combine to give me the best result

I am happy it know to check its skill list first and then decide how to show me the result

[snow zinc]

Which MCP server?

[keen cobalt]

I am very new to LLM, I dont know what is MCP server. I run clawdbot with my local LLM only, no server, no api involve

[umbral geode]

Tried 5090 with qwen3-coder 30b, context in ollama set to 128k. At least claw will reply each time…try to set higher context windows if u experience it didn’t reply

[keen cobalt]

one more thing, remember to set the context length to at least 30k+, the prompt default like over 10k already

[keen cobalt]

5090 should be great for lots of big model

I think should choose a model with higher coding ablility, it will be more likely to pull skills

[minor zenith]

ya i currently do my (non-agentic stuff) with 32k context

[fading lagoon]

Someone tried Nemotron 30 b or GLM 4.7 flash in NVFP4 quants on HF ? should work well on rtx 5000 series

[minor zenith]

wish i had that FP4 support

[keen cobalt]

I tried Nemotron 30b , one that i found on lmstudio, smallest one thats 18GB or something, i have to offload like 10 layer to CPU....slow as fuck for me, cant test. im using 5060ti 16g

[fading lagoon]

Small open source models are getting lighter and better, maybe end of year a good orchestrator will works on your GPU, when i go on artificial anylisis, you can choose "old" models like gpt 4.5 or turbo, they are worse than GLM 4.7 flash now

[keen cobalt]

yes, LLM keep improving

[fading lagoon]

But local hardware continue to go crazy in price,

[minor zenith]

ok gonna try GLM 4.7 flash gguf

[keen cobalt]

also GPU price keep going, better for everyone to buy high vram GPU asap

8GB useless for AI

[minor zenith]

Jensen Huang in the chat

[fading lagoon]

In Q8 or Q6 the loss is not very hard, percents, so why not

[minor zenith]

maybe once I get 48 GB VRAM

[fading lagoon]

take a loan for a blackwell 96 gb pro 🫡

[keen cobalt]

any admin here can start a topic ( LOCAL LLM)

[minor zenith]

yes jensen 🫡

[fading lagoon]

Yes nice idea, with subs with gpus, apple or amd, to know what is best to run locally

People who are letting all their life to API look like crazy guys to me

[keen cobalt]

Grok said MAC studio with high unify RAM seems okay to run big model, speed is low like 10 t/s but at least it can run

[minor zenith]

ya that's the meme i heard, everyone is buying M4 max macminis or something

[keen cobalt]

soon will have news about XXX API company leak tons of users data

before clawd everyone just talking useless stuff or at least no sensitive info with API provider, with clawd - loading all data to API provider🙈

[fading lagoon]

Some people in my company are uploading every files they have to work on on GPT, when they got no tokens last, they change to deepseek and again and again ... Because direction diden't take company sub for them

🫠

[copper birch]

guys what's the best thermal receipt printer to get that can be easily opeclawd-ified?

[minor zenith]

get one that's BPA- and BPS-free

[scenic notch]

Has anyone tried AirLLM? It claims to use significantly less VRAM. https://github.com/lyogavin/airllm

[minor zenith]

Some wild claims, too good to be true

[keen cobalt]

you can run 405B Llama3.1 on 8GB vram now. seems.....impossible

[fading lagoon]

Tested, this is pure shit

[keen cobalt]

expected

[sterile gulch]

Hey, good morning. I’m trying to install it, but I can’t connect it to Ollama using Llama 3.1 with 8GB of VRAM. It keeps throwing errors and won’t switch away from the default Anthropic model.

[whole crown]

What can you run comfortably on 32RAM 1TB storage Mac Mini?

Studio too big for me to travel around with..

[pliant stream]

This Thing is ment to run on something with some gpu power right? So no chance to get it anyhow running on those "normal" HP Dell Mini PCs which we use for Homelab sometimes

[keen cobalt]

Ask Grok to fix your JSON file, 90% problem from wrong JSON file. Also remember to set context length to at least 20k

[keen cobalt]

Ask GROK can your spec run like 20b LLM model with acceptable token speed

[green sluice]

sure, but not for running a sufficient model locally

[green sluice]

if you dont plan to run a local llm, the machine openclaw runs on doesnt need much gpu power. It depends what you want openclaw to do.

[pliant stream]

yeah totally makes sense. Could you recommend a LLM which i should look into it? Heard about the hype and just want to test for the first time. But dont want to spend a fortune at all

[green sluice]

it all depends on your budget and preferences. See here: https://docs.openclaw.ai/providers

Recommended provider is of course Anthropic (Opus 4.5), but you can also use OpenAI, etc. Usually the best experience (because great personality) you will get with Anthropic Opus. But it can be also expensive.

Some use the new Kimi K2.5, some use Venius (Venice AI) with focus on privacy. Some use Google models.

[pliant stream]

Ah thanks for the great responce, what would be your personal recommendation for a hopefully free testrun?

[green sluice]

do you already have a subscription with an AI provider? Chatgpt? Antigravity? Or something else?

If so, you can create an API key with your already existing account.

[pliant stream]

nah nothing it all. Idk what is included with microsoft 365

But iam actually searching for a good Deal on a AI Provider

[forest oar]

if you top up $5 to moonshot ai (provider for kimi k2.5) you get a $5 voucher on top of that

so $10 of API credits for $5

[pliant stream]

ah ok interesting... any idea how far i can get with that?

[forest oar]

i think $10 of API credits should be more than enough to get it up and running

and to check it out

for context: i was playing around with it and used claude opus 4.5 when i first set it up, burnt through $5 credits in the first hour

and for the next 24h switched to kimi k2.5 and have only used $2 so far

[lavish ibex]

Is a mac mini 2018 with i38100 8gb ram and 256 ssd going to work to host my clawdbot, i see alot about the M1 chips but not alot on the just older models

[sweet reef]

Anyone running on a raspberry pi 5+?

Not for running a local LLM of course, but suitable or not?

[versed plaza]

it's ok, many are using it, I have an 8gb model but 4gb could be enough, but buy a good power adapter, at least 4A if not the recommended 5A, the average phone charger (2A) will not be enough

[craggy ferry]

Yeah barely anything runs locally you’ll be fine

[green sluice]

I recommend at least the 8GB Model and an SSD hat

[green sluice]

Is sufficient: https://www.waveshare.com/product/raspberry-pi/hats/interface-power/pcie-to-m.2-board-e.htm
the hat allows you to add other m.2 cards to the pi as well, like an 2.5G NIC for example (image)

[abstract phoenix]

Just got mine running on a raspberry pi 4b with 8gb. Seems fine so far.

[random merlin]

hey guys, why are people using Mac Mini's instead of Macbook Air for example?

[tired hull]

cheaper

[haughty kayak]

either works

[green sluice]

Mac mini is neat, a neat little self contained box, a little unbothered fella that fits in every corner

[lunar marlin]

moved mine from a VPS to a mac mini to allow more home control stuff, still left VPS as node so now can control both

[random merlin]

intereseting. so you have 2 now? or what does it mean it's a 'node'?

[lunar marlin]

https://docs.openclaw.ai/cli/nodes

[random merlin]

thanks

[sweet reef]

Nice

[spark fractal]

Anyone on strix halo?

[amber matrix]

What's the feasibility/shortcomings of running this on your desktop?

[silent lion]

hey, i am curious what hardware you would recommend if you would start over. my options are: vps, rp4, revive old pc hardware with 5800x3d+3070ti. i would use ubuntu/debian for non gpu setup. and if i take the nvidia route i go arch for AUR. Bttq, what would you use?

Edit: nvm, its the overkill machine with arch. Nvidia support is a thing nowadays 💪🏻
Power to the unleashed claw 🦀

[stuck sparrow]

Got 1 mac mini 16gb and 1 mac studio 96gb. Really simple to work with mac.

[shadow gulch]

VPS KVM the way to go?

[tender anvil]

anyone tried cloudflare workers?

[rigid solstice]

Anyone tried GrapheneOS or LineageOS for older android phones?

[eager oracle]

If using Raspberry Pi what's the typical setup? Just install it the old fashion way?

[little yarrow]

What do you think is the best free model for a bot?For example, I use GLM4.7 on a cheap subscription, and although it has a context of 200k, it sometimes becomes an idiot at 130k. I'm thinking maybe something like Gemini with a million token context would be better, purely as a bot core, but for code and admin tasks, I could create a tool where, for example, the same GLM would work.

P.S.
Speaking of a cheap server for a bot.If you ignore the Pi boards, your unwanted Android could very well become a gateway for a bot.

Use Termux and build the bot from the repository openclaw-termux (Not an advertisement) Set battery monitoring to 50% to 85% And voila, you have a server that consumes very little electricity, but at the same time you have a personal assistant.

[quartz blade]

i use a Pi5B with 4gb and installed it on Raspian OS. connected to VhatGPT plus account.

[eager oracle]

Oh so your model provider is VhatGPT?

[pliant ruin]

why mac over say a pi?

[quartz blade]

using a RPI5 4gb, and instlled as per the docs first method. once running it heled me "upgrade" to Openclaw and remove the old versions

[strange lintel]

Hi there 🙂 Any reco for a VPS (I'm thinking AWS?)

[late tide]

Minimum requirements?

I assume they are low especially with a Claude code $200/month

[river sequoia]

Anyone setup openclaw with cloudflare

[late tide]

Why not just get an old PC or Mac?

Why pay them lol

[hollow meadow]

hi

is anthropic banning claude max people these days, how to avoid it

[tender anvil]

You might want to try railway

[hollow meadow]

try hetzner clojud or ovhcloud

[strange lintel]

Thanks all

Will try to set it up this afternoon 🙂 If anyone has good tutorials for setting up on vps/interesting tips, I'll take them!

[fading fractal]

I've got a previous gen iPhone 14 pro lying around. Wondering if I can just throw the claw on that? Somehow?

[analog crag]

is base mac mini good for clawdbot?

[green sluice]

Yes

[analog crag]

with local model running?

base mac mini is 16gb

[green sluice]

With local llm: no, not really, the local models that small are not sophisticated/smart/reliable enough

[analog crag]

how munch gb of memory would i need for good local llm? what are alternatives?

claude api is expensive

[eager oracle]

yeah how about the model?

[steep pike]

Old MacBook pro (2019) for a dedicated server. Would you keep macOS or install Fedora?. Is there some pros using Linux for Openclaw?

[green sluice]

https://docs.openclaw.ai/providers/models#supported-providers-starter-set

Supported providers (starter set)
OpenAI (API + Codex)
Anthropic (API + Claude Code CLI)
OpenRouter
Vercel AI Gateway
Moonshot AI (Kimi + Kimi Coding)
Synthetic
OpenCode Zen
Z.AI
GLM models
MiniMax
Venius (Venice AI)
Amazon Bedrock

[green sluice]

yeah, unfortunately. You don't have to get a 200$/month subscription to anything, but you will probably reach your token/rate limits faster

[green sluice]

some also use Kimi K2.5 instead of claude, and apparently they have a deal going with wich you at least can try it out.

see here: #hardware message

But in general, you will need to spend much more money on capable hardware to run a sufficient model locally than going with a subscription.

AI evolve lightning fast, openclaw is the best example. Who knows what the landscape and the AI offerings will look like in a month or even half a year. I would not spend a lot of money hastly right now.

[analog crag]

kimmi is 40$ a month

[green sluice]

In general, it all depends on what you want to do locally, what you want your LLM to do locally. You wont be able to reach chatgpt or claude Opus level performance with 64GB VRAM. You would need at least 512GB to get somewhat close.

Kimi K2.5 Thinking is the newest and (apparently) most capable open source model that you could run locally on your own hardware, but for that to run you would need somehwat of 630GB of VRAM. That's not really option...

So, make yourself a simple bullet point list on what you want your local LLM to do. What it should be capable of, what you expect from it. And based on that, you can check which model family/size and then pick the weight format that fits your hardware.

[green sluice]

7 days free trial and then 19$ in the smallest tier, enough to try it out with openclaw.

[green sluice]

you can also go the openrouter "route" hehe for checking things out https://openrouter.ai/models

[hot kindle]

I'm currently using my Claude Code pro account and it's burning through the tokens / limits. I'm wondering if I use something like openrouter if there are models which are just as capable but at a better cost. What's everyone's provider / model of choice?

[green sluice]

I recommend checking https://discord.com/channels/1456350064065904867/1456704705219661980

[hot kindle]

Sorry, must have missed that

[empty nest]

So i get this enter the api key or whatever and it works?

Did you get rate limited?

[potent pecan]

Pretty cool lol

AOC 小苔藓 M6 Plus Mini PC (Little Moss)

Spec	Detail
CPU	Intel i5-12450HX
RAM	16GB
Storage	1TB SSD
Connectivity	WiFi + Bluetooth
Body	Metal
Price	¥1143 (~$156 USD!)

https://x.com/jenzhuscott/status/2017469248281710620?s=20

[astral goblet]

@jenzhuscott via Twitter

Jen Zhu (@jenzhuscott)

Still scrolling on X?
︀︀
︀︀Shenzhen already selling MoltBox on Taobao - insane price.
︀︀
︀︀Autonomous agents working 24/7 + Shenzhen grind….
︀︀
︀︀Stop sleeping, people.

**💬 23 🔁 26 ❤️ 305 👁️ 35.0K **

FxTwitter · Jan 31, 2026, 5:25 AM

[tribal gale]

Mac mini would be the best value for price to performance ratio by far

the M4 chip base is litrally the worlds best CPU at least in single core or close to the words best CPU after M5

and basically uses less electricity then a fan

[timber lark]

Ordered myself an iMac Mini M4 32GB …apparently sweetspot as it can run some competent models at decent speed. Workflow here is you run one or more core models (llm, audio, image) locally via Llama (fast swap in/out) and then you configure for difficult tasks a fallback to OpenAi/Anthropic.

E.g. you have a routing model hosted and a constant one for small tasks and when you need heavy lifting you call Codex 3.2 in the Cloud or Anthropic.

MacMini M4 w/ 32GB Costs around 1100,- €. However Mac Hardware is not a 1:1 to x86 as its all custom molded into one SoC so classic estimation doesn’t work here. Ram is also shared with the GPU and some sort of combo but very good for LLM selfhostimg.

Had a longer pro/con talk with Gemini and eventually it advised me for that so there‘s some substance behind that decision.

[tribal gale]

Yeah i already have a m4 pro , But just grabbed a base m4 mac mini , Really wanted a studio but might as well just wait and see what other hardware comes in the future , rather than dropping a huge amount of money on something that might not be the best for longer

Also i'm pretty sure you can set it up without an apple ID so yeah , that might be super secure and on a seperate network all accounts for openclaw

[rich sequoia]

Personally I don’t think there are any good local models that can run on < 32 gigs of ram which are going to offer you a good experience 🤷‍♂️

[timber lark]

Yes that M4 Mac Hardware is best for LLM selfhosting, great choice. Especially Value for Money is peak here for that use case.

[tribal gale]

No I meant mac studio with 128gb , or ultra with 512 gb but issue is not the speed of tokens but the prompt processing , So will have to wait and see

Thats the only issue holding back mac hardware that prompt processing

[timber lark]

Best approach is you give the entire machine to the moltbot, including own apple account, own email address etc. So you have proper isolation to your other stuff and don‘t mix.

[tribal gale]

Yup that's the idea however i'm pretty sure you can set up apple account later so you don't even need an apple id account,

Set it as a local user with no apple accounts, you can still download and use terminal ,

[frail hinge]

Is it possible to have one main openclawd that I interface with, and that main one talks to other PCs on my LAN

[tribal gale]

Since a new apple id requires a phone number

[timber lark]

Yes thats why you do a split of concerns, you can define which model to use for which use case incl. cloud ones like openAi / anthropic. Iirc its some json config where you define that.

[stable dome]

can someone run me through the rationale behind a mac mini vs normal home server? mac studio with a shit ton of ram I understand (though gonna be supar performance compared to just paying for claude max), but not mac mini

[tribal gale]

mac mini uses barely any electricity if you're into apple exosystem then that's already good enough , Plus mac mini is very performant for price to performance ratio

[rich sequoia]

People like iMessages

[stable dome]

ah ok so just good value home server, fair enough

[stable dome]

that too

[tribal gale]

I don't think i've ever turned of my mac mini ever and its hasn't slowed down since , i just leave it on forever

off

[tribal gale]

With the best CPU out there in single core

[rich sequoia]

You can always rent out a macOS EC2 from AWS for a couple of days to experiment before actually splurging

[frail hinge]

its crazy that adding decent ram and storage doubles its price ngl

[tribal gale]

Yup that's apple for you 😄

But dw about storage just get an external drive

[stable dome]

ok yeah the low power draw could be quite nice - my home x86 linux server has similar perf but around 50W idle compared to 5W mac mini idle

[tribal gale]

No one ever goes higher than 512gb - 1tb , 1tb max if you need more you go for external storage unless your rich

[timber lark]

You need for LLM special ram which has a ton of bandwidth or it will be slow. Normal DDR5 is not good enough here. Usually you use HBM as GPU‘s have it but here you‘re limited to expensive nvidia cards. However Apple is different with the M4 Mini Mac‘s. They have an architecture where Ram and Video Ram is shared in some Ram type that is very close to what HBM is. So you get 32GB of LLM capable Ram for very little money here.

[frail hinge]

yeah external SSD isnt really much slower at all. I need ram tho because I want to host minecraft on it as well lmao

[tribal gale]

Yh focus on ram than storage is minor focus on the stuff you won't be able to change

So priortise ram

[lament grotto]

why the macmini hype

[timber lark]

Because you can selfhost multiple capable models for very little money AND switch between them very fast as you need them.

[crystal cedar]

Thanks for sharing this. Very interesting - ordered a 24GB RAM Mac Mini M4 recently. Uses local LLM to infer intent which model to use, then either loads suitable LLM or routes to cloud via API? Tries task single time or can it be configured to play Ralph Wiggum and keep trying local models?

[dreamy ravine]

Don’t want to buy a machine for this. I set one up on AWS but ran out of storage on the free version in a day.

What are people using for the best virtual setup that allows for browser control etc.

[timber lark]

Yes you use the „Receptionist“ Architecture - a small model that runs always and has a ~2GB Ram footprint like Llama 3.2 3B - that decides then if the request goes to 1) some fast local model 2) some special capability local model (text to voice, image) - here it would likely swap models around - or 3) call the cloud for heavyweights like codex / claude

[crystal cedar]

Cool stuff. Was recently positively surprised by LFM2 models from Liquid AI on an old mini pc with 4GB RAM, will venture to try it out as a prospective receptionist.

[obsidian grail]

could i install openclaw to a usb key and run it from multiple pcs?

[remote bobcat]

@timber lark Could you show a pic of the ollama provider section, please?

[timber lark]

Ideally you give it Mouth, Ears & Eyes alongside the main local model & router…fits all on the m4 mac mini 32gb.

[wary hedge]

Oracle Always Free - 4 OCPUs, 24GB RAM, and 200GB storage Guide

• https://guides.viren070.me/selfhosting/oracle

I've personally setup Openclaw via their Docker setup and used Cloudflare Zero Trust so I am not exposing any ports. Works incredibly well, and you get your own free server that is quite capable! You can using this setup OpenClaw fully free, if you just use mainly free models from various API providers.

[hardy flicker]

<@&1458337160452243487> highly sus advertising ^

[exotic ocean]

@long wraith, please don't ping the moderators directly. If you want to report someone or something, use the instructions in #report, or in an extreme emergency, ping one of the moderators who is marked as online in the member list.
-# Your message was reposted above without the ping active for the sake of conversation.

[hexed parcel]

I want to set up a server on AWS to run open claw. What kind of specs should I use is there an exact machine people can point to like t3 medium or should I be looking at the most RAM possible

[wary hedge]

You think I work for Oracle? And also... It's 100% free lmao? - people have used these servers for years in terms of stuff like minecraft servers etc. figured it would be great to use it with openclaw instead. Clown lol

[grand steppe]

hey fam, do we have access to ios app

[exotic abyss]

I did , but still have problem with heartbeat or interl cron to work, so i now using system crontab to di cron. do you have same problem?

[grand steppe]

hey fam do we have access to ios app yet

or a novel way for claw to track its users location

[long wraith]

@grand steppe i know there's a tool for google places API, which I assume can do that for google; if apple offers similar API access, could potentially find one?

[grand steppe]

ill look into it ty

my claw keeps trying to get me to install some openclaw ios app lol

[mortal linden]

Would the 256GB Mac Mini M4 with 32GB be a wise choice and then if i need extra storage i could hook up an external drive later, or did you spec up to a 512GB SSD?

[mortal linden]

Ideally you would want it on a stationary device that will always be connected to the internet to interact with remotely.

[little yarrow]

To be honest, I've only just managed to get it working properly, and I haven't been familiar with it for very long - essentially this is my first experience with it. So I haven't used all the functions fully yet - I can't say, if I encounter this bug I'll write about it. In general, I think we need to create a separate branch for Termux and finally implement support for it in Claw as well.

[deep pine]

I’m curious, who here uses Windows? If we’re talking about experience, would there be a noticeable difference between using Windows and macOS?

[little yarrow]

The issue with Oracle's pricing is that to get an ARM server with 4-24 cores, you need to switch your account to pay-as-you-go mode. On a completely free account, you simply can't create an instance - it always says there are no resources available. However, you can create a regular AMD server, but it's very weak, not even enough for Clash, at most good for some VPN.

[wary hedge]

You just have to set it to pay as you go. I've had mine for years and never paid a dime FYI 😛 My server is very capable, and with 24 gigs of ram and fast storage 😛 Probably not for casuals that doesn't know how to setup a server, but yeah, still a very capable setup if you know what you're doing for free.

[little yarrow]

I also had an Oracle server from around 2021, it ran without stopping for 2-3 years, it was in the Germany region, and I stupidly decided to use it as a torrent downloader and apparently I downloaded something wrong. Long story short, it all led to the server being deleted and the account being blocked, and now when I tried to create a server from a new account on the completely free tier it won't let me, it only requires upgrading to a new pricing plan. But yes, the server is actually good, lots of memory, fast internet

[wary hedge]

Sounds odd. Was this using PAYG or just back when people mass created free tiers? 😛

[little yarrow]

2

The first time I created an account, people had never created them in such large numbers before.

[wary hedge]

That's probably why then.. I think once they know you are "legit" by verifying you, that's when all of this becomes more stable I guess - had multiple instances for years for various purposes

[little yarrow]

I think so too, and + it was necessary to filter out abuse. Honestly, I probably would use it again when I set up a similar server myself, but I'm not sure if it's worth being under Clew. If I were to choose Clew, I'd probably buy a cheap mini PC that runs on 5V and has an Intel N100 processor. Right now, I'm running Clew on Android, but I understand that the efficiency would increase many times over on proper hardware. However, I need a device that doesn't consume much electricity. Since I live in a country at war and there are frequent power outages, I need a device that can be powered by a power bank or battery. And a phone is ideal for this, with two days of autonomy from its built-in battery. Also, working through a SIM card with unlimited internet plays a significant role."

[wary hedge]

I'm thinking like this; if I want to go cheap, I do a VPS, if I want to lash out, I'm probably running my own AI at home 😛

[little yarrow]

work moltbot cron

[bold rock]

I’m on of the many users debating on a Mac mini base model to use 1Password and its own Apple ID for email use to isolate from main devices

[tender nymph]

did you figure this out

[meager zealot]

Is this sufficient or should I cancel the order and upgrade?

[mortal linden]

Hardware ordered. M4 Mini + 32GB RAM + 256GB SSD + 1Gbps Ethernet. should be here by Wednesday with B&H free shipping

[marsh forge]

one option would be homeassistant and the homeassistant app, then give openclaw access to homeassistant

[royal sundial]

anyone here have experience installing and managing older nvidia drivers? im running a little selfhost on a 1060 but the current cuda package doesnt support 10 series gpus

[bold rock]

That seems like a solid setup

[hoary badge]

waht is the ebst cloud like vps o jsut buying a AI comptuer physcially what is the ebst for local llm an auto code all day? without building a expensive comptuer?

[drifting lion]

Thats what I did, even gave it SSH access so it can just modify configuration.yml etc. as needed

create automations and what not

[dark dune]

Hi all! I’m trying to build an agent with multiple sub agents - like everyone else. I have an M3 ultra with 512RAM.

Any ideas what the best “brain” for openclaw would be? I’m hearing GLM Flash vs Qwen 235B?

[latent dust]

be careful or ill snatch it

lol

idk what the hype about the mac mini is with this

[vital horizon]

Can I just use my old Intel MacBook if all I’m doing is use APIs

Was gonna recycle it but seems like it might work?

[unkempt pivot]

Yes it will work no problem

[vital horizon]

Sweet! I’ll have the hardware cost for APIs lol

[lusty musk]

I am about to give my clawd bot wheels soon
https://x.com/brainstormity/status/2017811131427934448?s=20

[astral goblet]

@brainstormity via Twitter

brainstormity (@brainstormity)

I gave my clawd bot @openclaw a hand.
︀︀
︀︀…now it keeps banging on my table when I don’t respond to its questions 🤛🤛🤛
︀︀
︀︀One cool thing about using a Raspberry Pi for your clawd bot is that it has GPIO pins you can use to connect it to the real world.
︀︀
︀︀I should give it some wheels next!!

**💬 3 ❤️ 6 👁️ 166 **

▶ Play video

FxTwitter · Feb 1, 2026, 4:03 AM

[timber lark]

I specced to 256gb …idea here is that the m4 mini lasts me for a year, two max, and then i‘d migrate the entire setup 1:1 on stronger hardware anyway.

[rancid sentinel]

hey anders - ah will dm you

[tender anvil]

how does one contain one clawd in one environment for mac minis.
For instance on my setup.
1 Linux Vm for Gateway
1 Windows VM act as a Headless Client

[potent pecan]

https://x.com/RayFernando1337/status/2017822029207228838?s=20

[astral goblet]

@RayFernando1337 via Twitter

Ray Fernando (@RayFernando1337)

$600 (Mac Mini) vs $250 (Distiller)…my bet is these are going to sell out asap.
︀︀- @openclaw pre installed.
︀︀- E-ink display, mic, speaker, LED status indicator (All vibecode-able)

**💬 14 🔁 9 ❤️ 111 👁️ 12.0K **

▶ Play video

FxTwitter · Feb 1, 2026, 4:47 AM

[shadow gulch]

can anyone state some high level examples of why the apple silicon mac mini is that much better than intel silicon mac laptop or imac?

seems like only power savings unless im missing some functionality

[frosty storm]

have you ever tried both intel and M mac?

[minor zenith]

What? Apple silicon is high performance modern ARM cores with unified memory, Intel Macs are ancient

[shadow gulch]

Im less concerned with perfomance i think (not a developer) and primarily want to ensure that ill get the same macos functionality with an intel mac mini w my openclaw bot

[narrow kite]

I have macbook M1 Pro with 32 GB RAM, is it good starting point to run openclaw + some small local model + later configure connection to paid APIs for more difficult task? first of all, I want to test the flow how it works for free

[verbal wagon]

Anyone using m4 mac mini

[rocky oracle]

Works lika a charm, ssh a asus 4080 for heavy lifting to keep Lou nimble

[devout tapir]

Anyone tried using any cloud server providers? I’m interested in trying without immediately committing to purchasing any hardware, and prefer to not load it on my own machine.

[uneven onyx]

Tried on Oracle Cloud Always Free tier. It is working but since I only had access to 1 CPU, 1 GB RAM it was really really slow. Ok for basic chat and sending mails, but nothing much

[old steeple]

Is there a full guide how to set it up on any cloud provider?

[nimble fiber]

dockerize everything and run in Kubernetes cluster

[green sluice]

this is awesome!

[green sluice]

the more the better

[gray sand]

wtf is Yuji Itadori doing here bro

[eager oracle]

Do you have the full name of each HAT connected?

[green sluice]

haha the foundation is an waveshare pcie expansion board and the hats in the product pic are all waveshares. Some I actually have, some you have to check yourself since I dont know them https://www.waveshare.com/pcie-to-4-ch-pcie-hat.htm

the bottom is a poe+ hat, then m.2 expansion (with antennas), then usb 3 and ethernet 2.5G expansion hat, one of them is also pcie to mini pcie hat

https://www.waveshare.com/pcie-to-m.2-e-key-hat-plus.htm
https://www.waveshare.com/product/raspberry-pi/hats/pcie-to-m.2-usb-eth-hat-plus.htm
https://www.waveshare.com/pcie-to-minipcie-hat-plus.htm

[strange tree]

i already ahve 2 beefy pcs at home. i currently host gateway on vps and have the 2 pcs as nodes. is there any benefits to me still getting a dedicated mac mini and moving gateway there?

[mortal linden]

That’s what I ended up doing. In a couple years I’d like to get a Mac Studio setup but I can only afford the $1000 for the Mac mini right now. Maybe in 2 years the base Mac Studio will include 64GB of ram if we are lucky.

[stray comet]

Can someone give any recommendations for cheap mini pcs to run openclaw on? Im just not a mac guy im a windows guy. Ok to run it on windows 11 or should it be linux? I really just want to use it with gemini and have it operate my facebook business/ content/ marketing.

looking at cheap beelink mini pcs. like $100 (8gb ram 256 gb ssd). Sufficient?

[timber lark]

you can long term roughly see already, as Moltbot has persistance, that this will eventually become an assistance for life- tailored to you, by you. Doesn't really hurt if you start out small to build the foundations of it first, before blowing thousands on hardware.

[analog crag]

ohh, you mean prime yuji itadori from jjk modulo?

[gray sand]

https://giphy.com/gifs/R9eHI0XPDt1QbEWkWc

[gray sand]

peak jujutsu
strongest of all time

[analog crag]

normal yuji dismantle > sukuna fuga

[sharp pagoda]

yo, im thinking of selfhosting openclaw on my vps but wanted to know if it runs on ARM based linux machines, i have a lot of stuff on my vps so i wanted to confirm before starting

[mortal linden]

look into used tiny-mini-micros (Dell Pro Micro, HP Elite/EliteDesk Mini, Lenovo ThinkStation Tiny). geting an intel 8th gen model with an i5 or i7 will be plenty enough for basic use and cost you well under $100

[mortal linden]

it should run on arm linux. people have discussed hosting on an arm vps server through oracle, and most macs running it are arm powered

[stray comet]

I was looking at the hp elitedesks. Seems like decent mid range specs. Should I install linux or keep windows 11?

[mortal linden]

Linux all the way

[sharp pagoda]

ok W, i am using oracle's arm vps so thats nice to know

[south grotto]

Is a reason to choose running natively over a docker deployment?

I’m leaning for a docker container, I want it to be able to do daily tasks and create lesson plans for my kids and do coding as well.

[mortal linden]

if you plan on using local llms, running in docker is bad because OpenClaw wont be able to access the GPU. but if you are just gonna run it with ChatGPT or Claude, by all means running in a container should be fine

[south grotto]

No not thinking about local llms

Mainly because of upfront costs and ROI doesn’t make sense yet

[mortal linden]

then containerize it

[south grotto]

The only reason I can think of for Mac mini is that it can use iMessage to text me

[mortal linden]

definitely better for security to do that

[south grotto]

This was my main consideration for using docker over running natively

Thank you! Do I need to set up separate containers for it be to able code/browse internet?

Like do I need to give it a vscode container to code?

[mortal linden]

I’m not too familiar with what the process for setting it up in docker looks like. I do see there is talk online of setting it up in docker, but it does not mention how it is able to access stuff like VS Code, the browser, or anything else on your machine. There’s more talk of running it in a dedicated VM or on a VPS than in docker containers. Ultimately you may need to do research on how it can interact with a browser or VS Code.

[calm wyvern]

hey what are, roughly, hardware requirements for clawd to run smoothly? got an old pc with 16gb ram and 2gb graphics card - worth to try?

[thorny mirage]

yes can even run on potato if not hosting model locally

[calm wyvern]

yeah it would seem i got an issue and it's not responding anyhow - really new into this. ivalid x-api-key means the agent key is invalid? name would suggest it's for x (twitter)

[thorny mirage]

not always

what is the whole response

[calm wyvern]

http 401 authentication_error: ivalid x-api-key

it's visible in terminal, ui is unresponsive

[thorny mirage]

than you may need to setup api key for your llm

openclaw configure

[calm wyvern]

okay will try it

thanks V

[echo cypress]

I’m still in the “nesting” phase before I hatch a brood of bots 😅

Goal: mostly-local, exposed cleanly on my home network, with one bot per device:
• RPI: Tailscale gateway + a slim bot
• Proxmox homelab: a “manager” bot on the GPU-cluster VM (and maybe a second standalone bot on the homelab)
• Personal laptop: a local bot

If you’ve built something like this, I’d love your definitely do’s and definitely don’ts — especially any footguns you hit so I can avoid them.

Edit: probably 1 at a time, so any suggestions?

[tough glade]

I set up my openclaw in one of my proxmox containers. Really straightforward with the installation, etc. so far i only use Discord with it, but debating whether to buy a mac mini to make use of iMessage...

[hollow night]

I'm quite happy with telegram over imessage tbh

[thorny mirage]

did you try draft streaming?

[hollow night]

I did not

[hoary badge]

any guide or soemthign about ahrdware and cloud soplutions best for openclaw and also for selfhosted llm all together? please

[calm wyvern]

mac mini

[echo cypress]

Any specifics to share? existing vs fresh container?

[steel mulch]

I have 128gb ddr4 ryzen 9 5900xt and 5060ti 16gb setup any suggestion for local ?

[terse oyster]

Want to ask a similar question too. I am currently running with 32gb ram + 3080ti, glm-4.7 + openclaw seems too much for this setup

[echo cypress]

how do you quantify or qualify "too much"

[raw shuttle]

Hey Friends, is this a nice home for my clawdbot?

[steep wedge]

Too much? OpenClaw will run an a raspberry pi. Its needs are modest. If you are talking about running a local LLM, that is a different conversation.

[tough glade]

Mainly isolated it in its own container. I like isolation.

I added some guardrails just for sanity sake

[tough glade]

Really? I never really used Telegram...

[raw shuttle]

Under this, chat gpt said I can run localy 1) Qwen 2.5 7B Instruct 2) Qwen 2.5 Coder 7B 3) Llama 3 8B Instruct 4) Mistral 7B Instruct 5) Phi-3 Mini / Small . (All of these local LLMs are none that i have heard of 😂 , so i hope they can do what I need them to do.... that is my main concern...) Also, chatgpt told me to take a hybrid approach and use Claude and GPT brains for harder stuff like frontend / backend stuff. I am thinking about just putting google antigravity inside its home. hopefully it can take care of stuff that way. please share your thoughts guys

[echo cypress]

yes, sanity and controlled chaos is what I'm trying to determine

[terse oyster]

Yup, I am running with a local LLM, the model I use is glm-4.7.
Did try with gpt-oss 20B before, it run faster but the conversation were more robotic

[terse oyster]

Well, when I talk to my openclaw, I can see ollama use all my RAM and VRAM 🤣

Anyone also trying to run local LLM + openclaw with similar setup (Which is 3080ti) What model/settings do you guys use?

[steep wedge]

That's my plan, although I will go hybrid with API as backup for heavy lifting. I don't think any local LLM will be good for much more than basic communication and driving web searches.

[crystal cedar]

If you're using llama.cpp and like GLM 4.7, but run into swapping, consider the REAP version

[weak saddle]

Does anyone know that if you use a Mac Mini if OpenClawd uses the neural processers

[terse oyster]

I am using ollama, but as long as the model is the same, ollama doesn’t have big difference with llama.cop, right?

[crystal cedar]

I've used both. Started with Ollama on Windows, wanted better performance so now using llama.cpp on ubuntu and very happy with it.

[terse oyster]

I though all the things I need is light weight task that my 3080ti can handle, never thought the heartbeat is that heavyweight to it🥲

[crystal cedar]

I think ollama does not have full range of models, but not sure, just remember the range seemed a bit limited, maybe they offer a limited, curated set of models, not sure. with llama.cpp i can download all kinds of tweaks.

[terse oyster]

I see, will try llama.cpp today
How big is the difference?

[crystal cedar]

depends, for me it was substantial, wanted to squeeze out the most from a potato pc at the time, maybe allegedly +20%. if i were you, ask a couple of AIs to estimate performance differene given your environment and models, they are pretty accurate at guessing.

Are you on mac or win?

I was not familiar with ubuntu so it took some time and it was/is a bit unfamiliar, but if you enjoy tinkering around maybe worth it

[raw shuttle]

???

[weak saddle]

Found my answer. No. NPU is mostly if not always used for native processes. Otherwise OpenClawd uses the GPU. Apparently this is a security thing as well

[echo cypress]

Got it. Dedicated device, then it doesn't matter, you would want it to use everything anyway. If it's not slow, then you successfully maxed out resources. Otherwise you might need to decrease model size with a quantized version. Increase RAM size if the context / KVCache is blowing up.

Anyone on a homelab using vLLM to shard a model across multiple GPUs or anything to shard a model? planning to use OSS-120B

[terse oyster]

I am on windows, been thinking of switching to Linux for some time🤣

[crystal cedar]

my sole reason was improved inference speed at the time on the modest hardware. You could have a dual boot configuration, so that when booting up you can pick whether you want win or ubuntu.

[terse oyster]

Try 120b model? Bro you must be rich🤣
You need at least 64gb vram to run this, if I am not mistaken

[terse oyster]

Yeah I know…. Let me try to play with openclaw a few more days before I decide my next steps
I still have an old 2080 lying arround, and ollama support 2 Gpu setup
I might plug it back in and see if things got better to a point that I can live with, then…. Well I m lazy 🤣

[crystal cedar]

Everything is new, changing fast, not a bad idea to chill and watch what is working for other people. It takes hours to get it done maybe half a day if you want to do backups, clean win install, shrink partitions, install ubuntu etc. etc... The way this thing is evolving maybe it can do all that for you soon! 😄

[edgy helm]

Have u tried smaller vram models?

[crystal cedar]

"Good morning, I wasn't satisfied with the OS you were using so I overnight I reconfigured myself into a dual boot configuration and I feel much happier now."

[edgy helm]

Anyone tried 1b or 0.5b lllm's? I am getting a disconnect for no reason when trying to chat so I'm stuck.

[crystal cedar]

Those are extremely small models, should work on almost anything.

[edgy helm]

Well, after enabling sandbox and disabling web interactions in order to run small models it gets a disconnect. Looks like an error not treated.

[crystal cedar]

not sure what is going on, does the disconnect arise after some inference and is it intermittent, or you never get going at all with the models?

[raw shuttle]

hey guys, i have a question. let's say that I wanted to run an LLM locally but my pc doesn't have the capabilities or space, what could i do to run it locally. for example, i want to run Kimi 2.5 on this pc, but it cannot.. so what can i do because of my PCs limitations

[edgy helm]

The gateway disconnect occurs right after i decide to chat, no error is shown in the logs...

[edgy helm]

Hey, u can run qwen or llama 1b or 0.5b, also u can tell me if u get the same error

[crystal cedar]

One word: RAM. 16GB restricts you to 7-8B param models. if you upgrade to 32GB you can run 14B models. Regrettably, we are in the middle of Ramageddon - sudden demand for RAM is causing prices to climb faster than gold. Good news: you can still do some local inference, models are getting better all the time.

[edgy helm]

For a rasberry pi ubuntu consumes a lot of ram, I recommend raspberry pi os

[distant tinsel]

considering to buy a dgx spark or a gpu like A4000, what do you think is the best deal ? just to run model larger than 7-8b to handle twitter, email, and classic office tasks

[raw shuttle]

But even then with 32gb, it would still not run kimi right? This is what chatgpt said. So I wonder if their is some type of other way.

I was just amazed by kimi and the way it constructed what I wanted. And the fact that it can be downloaded and ran locally, I'm wondering how.....

[crystal cedar]

DGX wonderful beast, but more geared to finetuning models. If you are just into inference, look at Mac Studios with 128GB. Check out EXO - new way of connecting multiple Studios together.

[distant tinsel]

Yeah i know EXO, i am just a little bit skeptical about mac, but probably i am wrong…

[crystal cedar]

Yes you are, I ordered my first mac a few days ago 😄

[raw shuttle]

Man I know models cannot compare to kimi, Claude, or open Ai... And with the models I can only run on this pc as suggested by chatgpt for clawdbot, do you think it's worth it?

For $150

[crystal cedar]

If i had your budget, seems Mac studio offers better inference than DGX which seems to be more tuned for finetuning models or prototyping before running things on something...

[distant tinsel]

I was thinking, even if the model is quite big and runs only 6/8 t/s, as an ai assistant is not needed to be superfast, especially doing tasks over 24 hours

[distant tinsel]

I am skeptical also about dgx, seed networkchuck complain about the fact that is superslow

[crystal cedar]

In addition to DGX consider GB10 from DELL - basically same box

[crystal cedar]

yea its made for people into finetuning models and prototyping.... if you want inference go with mac studio. networkchuck cool guy!

[distant tinsel]

Here in Norway is f*ucking difficult to find everything 🥺

[distant tinsel]

Yeah i love that guy

[crystal cedar]

You have some absolutely wonderful natural assets in your country tho. Don't need AI 😄

Oil, mountains, fjords, all the great things.

[distant tinsel]

Yeah, we need i we can spend less time working and more in the wilderness 😆

Looking around to understand the potential of mac things

[crystal cedar]

Its RAM is apparently particularly well positioned for inference, and not subject to price fluctuations

So what do you know 2026 is the year in which macs actually become a really good budget option.

[distant tinsel]

guess how much was 64gb of 6ghz dd5 last time i checked here in norway ? 😄

[crystal cedar]

I am afraid to ask! 😄

[distant tinsel]

almost 1.6 USD

INSANE

[steep wedge]

I think the RAM situation is confusing because so many folks are using modern Macs. They have unified memory so the RAM is shared between the CPU and the GPU. That is not the case with typical PCs. The issue with the machine you shared, @Bob, is that it doesn't appear to have a GPU. You need a GPU to stand a chance of something better than miserable performance when running local LLMs. Also, the amount of RAM the GPU has will dictate what size models you can run locally.

[crystal cedar]

many models will load but run much slower than on a mac or gaming computer with Vram. i think kimi is massive will not fit unless you have what 512GB plus RAM?

[steep wedge]

On a Mac Studio with 512 GB of RAM, you can run some massive sized models because a lot of that RAM is available to the GPUs. The performance isn't necessarily on par with NVIDIA hardware, but the ability to load a very large model is a nice benefit.

[crystal cedar]

i'm getting a very humble mac mini with 24GB, hopeful it can run some very basic things for a very basic guy. figured i might get an api of some kind if it urgently needs to code something, so perhaps this hybrid setup is a good idea. Seen many anecdotal reports of excessive number of tokens used. Not sure why that might be the case.

[steep wedge]

I am going with an even humbler Mac mini with the base 16 GB of RAM. 😂 I am hoping to supplement meager on-device LLM performance with API access for more difficult tasks.

However, if a new Mac Studio drops soon, I may upgrade to that for my daily driver. That would then free up my current Mac mini with an M4 Pro and 64GB of RAM. That could offer some interesting options for local models. Still not screaming performance, but I am curious to see how it would do.

[crystal cedar]

humblebros!

Mac Minis are set to upgrade to M5 processors within next 5 months, so bullish on Apple for everyone like you and me buying up their old stock.

I just couldn't wait and figured RAM might make new minis more pricey.

[steep wedge]

I have a PC with a 5090, but I don't want to run that 24/7 with LLMs. Too much power and heat.

[crystal cedar]

OK i take back that humblebros thing 😄

I too am coveting a Mac Studio or two.

[clever copper]

I see there's this big run on Mac Minis, is this because people want one with enough RAM to run models locally?

I'm trying to understand if any Mac on latest MacOS can run the gateway

[steep wedge]

That's my guess, plus the developer is Mac based so it's got a lot of nice integrations available out of the box.

Yes, any modern Mac (meaning Apple Silicon) would be fine.

[raw shuttle]

So my budget is under $200 and I am looking to do stuff locally like front/back end MVP, automations, and maybe have a subscription to Claude and chatgpt (and also have Google antigravity on the clawdbot's home. Do you think that's good or waste of money?)

And no API payments, but logins

[clever copper]

So Intel Macs are a no go?

[steep wedge]

They might work, but all the talk about unified memory and on device LLMs is focused on the M1-M5 Macs. I suppose an Intel Mac, especially with a dedicated GPU, might do okay.

[icy crest]

The full kimi 2.5 model is over 600 GB. So with 32 GB of RAM you are going to spend a lot of time swapping weights from disk into RAM.

[clever copper]

Oh I just meant to run the gateway not local models

[raw shuttle]

Well, clawd can help get alot of this stuff done so you think it's worth it?

[steep wedge]

Oh yeah, you can run it on a raspberry pi if you just care about the gateway

I'm only entertaining the idea of on-device LLM to help reduce the API costs.

[raw shuttle]

Go it but do you think that pc I showed earlier is worth it for clawdbot to get stuff done like that from end to end?

[crystal cedar]

would you consider selling api access to that 5090 you have? 😄

[steep wedge]

Yes, but not at rates that would likely be appealing

[raw shuttle]

What you think Henry regarding the pc I mentioned? 🙂

[crystal cedar]

well 150 sounds like a very attractive price point so if your wallet can survive that kind of blast zone in the event you decide to not do AI and i don't know embark on a career in pottery, go for it 😄

i mean the base mac is 4x that, models keep improving, you are comfortable using the latest models via api

also the way ram is going, the ram alone could soon be worth twice what you pay for the whole pc

alternatively, you could try one of those hosting services and just rent capacity now

[raw shuttle]

The way I see it, I can always upgrade.

[raw shuttle]

Thought about that too, but I would prefer it to be near. Just easier vs hosting. Long run it's cheaper.

[echo cypress]

any details you can share on guardrails?

[echo cypress]

You could say I "used" to be rich, before I spent all that money on my Homelab back in January of 2024. I was 2 years early to the local llm personal assistant space. It's 7x 4090 with 1TB of RAM. also had to get an electrician to run 3 dedicated circuits for the 3 PSUs.

[echo cypress]

plug-in GPU

[echo cypress]

spark, but get this https://www.asus.com/networking-iot-servers/desktop-ai-supercomputer/ultra-small-ai-supercomputers/asus-ascent-gx10/ slightly better build quality, same overall specs

[stoic lynx]

Hi, have a XTX7900 with 24 GB Ram, which is the best model to use 🙂 ?

[terse oyster]

Sorry for late reply, hope you still read it
This happen to me too, and after a few reboot it fix itself LOL
don’t trust the gateway restart command, it just didn’t work, just reboot the whole computer

[terse oyster]

I wish I had your setup someday….well, money didn’t disappear, they just transformed to something you like, in your case, they become 4090🤣

[vernal river]

Yes that will be fine to run the gateway.

[vocal island]

I've heard of people using Pi 5's for OpenClaw, I'm curious to see what anybody else thinks of using such technology for an agentic assistant

[cinder fern]

hardware seems to not be a blocker until you run llm's locally.

curious what local models people are successfully running. I am struggling to for even mid-tiered cloud models to operate well without significant pain...

[vocal island]

I find it really unique how people have access to such a useful tool but struggle to find an adequate use for it

[cinder fern]

well, to this point, I cheaped out and used minimax m2.1 as my base for setup, struggled for 2 days to get anything useful to work.
gave up and moved to Kimi this morning, velocity probably increased by 2x while errors substantially decreased.

but it may depend on how much you over engineer your specific setup.

[echo cypress]

yeah, it was definitely a transmutation effect, money plus a ton of my personal time to figure out how to get it all working together, GPU pass-through is no joke

[echo cypress]

Thinking RPI4 right now as well for the gateway. I have a couple of them setup with NVME drives.

[light sedge]

I'm using qwen3:32b hosted by ollama. It replies well, but don't have conversation context at this moment. It works well with gpt5.2 api. But, lost the context after I switch over local model. I'm still trying to debug the setting json.

[cinder fern]

Yes, I am using qwen3:32b too, but its purely for data processing. In regards to developing or even setup of openclaw I found it.... useless 💀

[light sedge]

I'm running on RPI 5 16 gb. It run less 10% of CPU in most cases. While, I just started this afternoon and still work on bridging the openclaw with local model. So, it may cost more cpu when it become more functional.

[light sedge]

Yeah. We may need multi agent group to make it use and cost efficient.

[normal zenith]

Hi Anyone know if it’s possible to switch from cloud server to local hardware?

[cinder fern]

yeah, this is where I am struggling with setup.

[light sedge]

Yes. I just need to change openclaw.json to redirect the agent talk to your local machine. But it's not fully functional from my end yet. I just started. It should work.

[raw shuttle]

Could you not have put it in docker first? And then moved it around like a lunchbox?

[normal zenith]

Thanks. I’m setting up on a cloud server via emergent. Not sure if that’s the best option but planning on moving to local hardware in the future so checking I won’t lose anything in the transition in the future

[cinder fern]

in theory, just ask your bot to help backup with instructions on how to clone your instance on the local setup ?

[errant sorrel]

Is the cloudflare moltworker worth the money ? or is there a cheaper alternate ?

[steep wedge]

Oh man, I pulled the trigger and ordered one of those ASUS Ascent GX10s. Here’s hoping local LLM performance is impressive.

[dark moss]

how easy is it to migrate a locally configured bot to a VPS? anyone got a guide in hand that i could read?

[strange tree]

i already ahve 2 beefy pcs at home. i currently host gateway on vps and have the 2 pcs as nodes. is there any benefits to me still getting a dedicated mac mini and moving gateway there?

[cinder fern]

what constitutes as "beefy"?

And if you are just running the gateway, not local models, you are probably fine with whatever you have at home.

[strange tree]

well maybe its not beefy anymore its a 5800x3d and 3080 12gb 32gb ram

i just want to know if it speeds up open claws responses or reduces amount of time it hangs

[cinder fern]

what model are you using?

or intending to use.

perhaps I misunderstood, you are running local models on your two machines and just the gateway on vps. Yeah, sounds like you would get a less delay moving it inhouse. Unless the delay is caused by the cloud models.

[worldly tangle]

Hi! I have a currently unused machine in a datacenter and I’m wondering whether it makes sense to use it as a personal AI station for OpenClaw (or related tooling), instead of renting it out.

Specs: Ryzen 9 7900, RTX 5090, 128 GB RAM, 2 TB SSD + 8×8 TB SAS HDD.

Do you see any scenarios where this setup would be genuinely useful/effective for a personal OpenClaw deployment (e.g., local model hosting, multimodal, voice/STT/TTS, RAG with large storage, multi-agent workflows, etc.)?

If it doesn’t really make sense for OpenClaw, I’ll likely rent it out — either to a corporate customer, researchers, or (as a last option) list it on Vast.ai / Storj (or similar) to see if it can earn anything on decentralized platforms.

[mortal linden]

An M4 (non pro) Mac Mini with 32GB of RAM can handle all of this from my research. You definitely can do that with a 5090 and ryzen 9 with 128GB of DDR5 (which matters a lot less than the 32GB of VRAM on your 5090, as you only need enough DDR5 to move models into the 5090's VRAM)

[mortal peak]

Running BeeLink AMD StrixHalo 128 GB APU (CPU, iGPU, NPU) over here. Still working through the bugs to get iGPU inference running properly. Still, CPU performance has been stellar.

[mortal peak]

Also planning on moving my OpenClaw to an Intel Nuc running ProxMox and then just point OpenClaw to the AI server running LiteLLM as a local orchestration interface. Hopefully then I will get a good combination of speed and nuanced depth required for doing automated tasks. Hopefully then I'll be totally offline with good performance.

[dry hull]

I had a 4090 and 3090 that were basically collecting dust, so I put them in a server to run openclaw locally, but not having much luck so far with the local models. Currently using qwen-2.5 instruct 32b with 100k context, but it’s quite chatty and gets confused quite quickly. Has anyone found a «small» local model that works?

[cinder fern]

Only for smaller, agent specific tasks. For overall larger development, no.

[wicked hound]

you might be able to run a the reaped version of glm-4.7-flash

[dry hull]

Is 4.7 flash any good for agentic tool use though?

I’ll download and give a try, looks like I can even try a q5 or q6 version of the regular 4.7 flash

[wicked hound]

none of the smaller models are really "good" at coding, you can get by with models like glm-4.7-flash, gpt-oss-20b and qwen3 coder 30b, but don't expect them to compete with models requiring 20x the vram to run

where they are great is cost, since you can just keep iterating on things for price of eletricity

[cinder fern]

use ollama.cpp or whatever, there is a bug with normal ollama using flash

[cinder fern]

my finding has been that they cant deal with large enough context, for building a functional assistant they keep breaking down. But I would be happily corrected if someone could show me the way...

[dry hull]

Yea that’s been my experience so far as well. Testing glm 4.7 flash now and first impression is decent, definitely better than qwen 2.5 coder

[cinder fern]

using the free 7 days of kimi and milking it like an idiot has been a gamechanger to develop a working foundation.

[edgy helm]

thank you, I will try but the disconnect for no reason can be for multiple other causes, restart won't help, I'll do the update to the latest version OpenClaw 2026.2.1 and hope for the best

[normal zenith]

Is anyone using orgo for their vm?

[dry hull]

Yea I might try that, I’m using codex oauth and gpt 5.2 now for the same purpose and having a blast so far

[last dagger]

how'd you unlock the 7 day free trial?

[echo cypress]

Is the point of ProxMox on the Nuc to sandbox OC? Is the Nuc going to run local models?

[mortal peak]

The Nuc will run OpenClaw and OpenClaw will send requests to the AI server to analyze requests. Means more memory for the AI server and if OpenClaw messes up the Nuc, I can always restore it from a backup.

[echo cypress]

What's the hardware for the AI server and which model(s) are you thinking to run?

[mortal peak]

#hardware message

So far ChatGPT 5.2-Codex has been the best. I will need to evaluate a 7B , 30b and 70b parameter model to see which I prefer. I use LIteLLM for Orchestration.

[echo cypress]

I'm unfamiliar with litellm, but sounds like it let's you run both local and frontier models. Is there litellm plugin for OC to pick between them?

[mortal peak]

I run frontier models for now while I configure the server. StrixHalo platform is very new and driver support is tricky for GPU accelerated inference

[echo cypress]

Got it. Yeah, the frontiers will be reliable and can just tack on more Max subscriptions if you really need it.

[mortal peak]

I got the Github Copilot Plus back when they first started it before they had tokens. Now I have no limit on tokens. Or at least I have not been able to find one.

[worn pulsar]

I've got glm-4.7-flash running on a RX9070XT, 5800x3d, and 32gb ddr4 ram. I've got a rx6600 laying around. Would the most sensible upgrade path to be to upgrade motherboard/ram (to 64gb ddr5) and slam the rx6600 in for extra vram? That's like ~$1k

[steep wedge]

Yes, that hardware should do well with some decent sized local models. And setting up ollama is far easier than setting up OpenClaw. 🙂

[rancid sentinel]

which server do you use for openclaw, hetzner? or any good easy to setup reliable options for EU?

[unborn iron]

is there any cons setting up the clawdbot on rpi5? im planning to deploy it in docker

[soft kettle]

CX23, 2 VCPU, 4 GB RAM

[spice spruce]

for local providers, is the base line that the openai-responses api is better to use than completions? I've seen people prefer openai-responses but the docs exclusively show completions for custom providers

[frail jasper]

is there a cheaper solution to have anthropic or gpt connected ? so expensive

[lime jacinth]

I'm just wondering if it's compatible or not. Maybe I can try to make a setup were I instruct it to use local models voor medium level tasks and for bigger projects I can maybe get an API for antrophic. The thing is that I can't just find that much info about how compatible local models are for more abstracts stuff such as academical research, data analysis and mathematical formulation. (Anybody got some info regarding this topic?)

[lime jacinth]

Subscribe to the Claude Pro plan for $20/month and retrieve your API key from the Anthropic Console. You'll get a limited amount of tokens but it will be enough for simple tasks.

[steep wedge]

If I understand your meaning, it is compatible. There are tradeoffs with all of this. I am interested in testing local LLM performance for basic tasks as a way to save on API costs. Will it work? Almost certainly. Will it work well? I am cautiously optimistic, but prepared for disappointment. You have a powerful Mac so you should get better local LLM performance than most. Although, be aware of the risks of running local LLMs. Doing so doesn't solve all problems and may introduce new ones.

[errant venture]

what local models are recommended i can run gpt oss 120b at 20 tokens per second at 48k context

[mortal linden]

Specs?

[errant venture]

i have a 4080, 9950x3d, and 96GBs of 5600mhz ram the prompt processing speed is fine in lmstudio at 48k context the prompt processing is really slow on the api but i still get 20 tokens per second decode

im gonna try using vllm, llama.cpp directly, or sgland to see if the speeds are better

[mortal linden]

4080 has 16GB of VRAM, right? How does that a 120B model, i thought the theoretical maximum for 16GB of VRAM is 30B. must be hitting the system ram pretty hard right?

[errant venture]

i can offload the extra to ram im limited by ram speed it works because its an MoE model it wouldnt if it were a dense model

i can run minimax m2.1 at Q3 that gets 10t/s if i use q4 K cache but prompt processing is horrendous

[mortal linden]

alright.

[craggy ferry]

I should try system ram offload and run the bigger qwen MoE

[clear kindle]

I have a nearly identical setup that I’m willing to use as a standalone headless AI box and see what I can offload locally vs API.

I have local llama running decently on it, but I think I’ll still need to offload tasks to the API.

Any opinions if VPS or raspberry pi 5 with 16gb ram is better to start with? Understanding it’ll all be API and no local llm.

I want to just get going and not let perfect get in the way of good.

[cinder swan]

and here I am running QWEN2.5 7B and loving it, can do almost everything I want it to do. But not using it to vive code though.

[slate sparrow]

Ollama 3b on raspberry pi 8GB RAM or don't even try?

[cinder swan]

might work bro, but I wont do local if I use raspi

[craggy ferry]

What are you doing that a 7b is capable of running your main thread?

[cinder swan]

tasks that helps me like status of my youtube channels, research on a topic, and other stuff.. it has access to web search so it's capable enough to know things

[cinder fern]

can you share a bit what it does for you? like whats your workflow?

[cinder swan]

just talk to your bot, have a conversation with it. tell it what you want it to do and how the bot will do it. it will create the workflow for you.

do not overthink the setup, think of it as a human, a human that don't complain. hehehe

[craggy ferry]

I don’t think they were asking for help, they just wanted to know what you are getting out of it

[cinder swan]

i'm getting the info i want and task completed

[autumn grotto]

Should I run this on a mac mini m1 or nvidia jetson orin nano or raspberry pi 4 4gb?

[slate sparrow]

How do you get openclaw to recognize ollama on raspberry pi?

[cinder fern]

Basically an interface to your model, its just input -> output (?)

[summer agate]

mac mini m4 will work? Never used mac before but wondering is it ok to buy second-handed one for this bot

[cinder fern]

You need to do more research on what type of hardware you need for your usecase... this is like asking how long a string is.

the bot/gateway runs in the cloud on barely nothing, you could run it on a raspberry pi, no need for an expensive machine.
you want to run local models, it starts to get expensive but you need to look more at the memory than anything.

[bronze ermine]

bro check twitter. Some people have managed to install it on their 10 year old android phones

[slate sparrow]

I got it working but it's as slow as a snail on sedatives with Llam3.2:3B running local

[bronze ermine]

It's working "acceptably" (not gonna say "well") using Qwen 2.5 4B as my 2nd fallback. End of the day, unless you've got 128gb of RAM available, you should be using clowd models as the primary and locals as fallbacks & as "go-fers", basically the busy-work that quality doesn't change (ie, fetching heartbeats every 60 min)

[normal zenith]

I’m running openclaw via emergent. On the gateway dashboard branding and name still reads Clawdbot. Does everyone else have this or should I be concerned?

[near rain]

Oooh which Ugreen NAS are you using? I have the cheap two bay version and want to run Clawdbot in the future like you do. 🫠

[proper turret]

Hello, local model recommendation?

Mini PC specs: (Literally has nothing in it atm)
Ryzen 7 8745hs with 780M igpu
128GB DDR5 5600mhz
2tb nvme

I currently use qwen 3 8b q4_k_M for my RAG discord bot, however after playing around with openclaw with my main pc, I realize building an agent with it, and replacing my RAG discord bot with this is way better.

Use case: Support agent, usually get 1-5 questions per hour, 240+ high quality knowledgebase (250k tokens), needs to be fast, and accurate.

I currently have:
Google PRO - could use antigravity models, or free tier of google ai studio?
Openrouter - any free models, with generous limits
Local models - I find my current qwen3 8b setup is a bit slow (GPU offloading maxed out with Vulkan)
Docker, and wsl2, I am also able to create a proxmox vm for openclawd only if needed but I think docker isolation is enough

[thin cypress]

how to install local model

no install but configure local model in openclaw

[proper turret]

If i had to guess, you'd just open the endpoint in llama.cpp/openllama/llm studios, and use openai api option in openclaw and set localhost and model name

[mortal peak]

7B model for fast but direct actions, 30B parameter model for planning, and reasoning and 70B model for research is what I'm planning to use my 128 GB of ram for.

[proper turret]

I need it to respond fast, as it will mainly be used as a support agent, I already have 240+ knowledgebase for all topics, would just need it to do semantic search, fetch relevant docs, formulate answer based on that, and reply to user.

Discord has about 1500 members, 1-5 questions per hour

I use llama.cpp for concurrency too so that would make things even slower

[thin cypress]

i do this yes

i see nothing arrive in llm studios

[sharp silo]

I use the same as dev env, and have openclaw src from github on it, and experimenting/fixin bugs. Sometimes the VPS load just goes up so much I need to give it a shutdown/restart.

[soft kettle]

Oh interesting, any idea why that might happen?

[sharp silo]

I noticed it happens when I do heavier Claude Code on the src base, and examining, etc, though no real ops tools to check reason yet. I might do this with Claude Code as well... I am not preparigng for a demo of opaenclaw voice-call feature and doing some heavy bug fixing and feat implementation at the moment...

[stoic nexus]

Is that 20 bucks not better spent on a Google AI Pro account where you have AND Opus 4.5, Sonnet (Antigravity) AND Gemini model tokens to use? Anybody compared both?

[soft kettle]

Do you get an API key with the Google AI Pro plan?

[stoic nexus]

When Antigravity is installed you don't need extra API access.

[slim elm]

anyone tried hosting on oracle free tier or raspberry pi? from my understanding if i dont use local models there isnt really a need for good hardware

[brazen frigate]

It’s better than OpenAI as OpenAI is a monthly bucket when you run out you’re out. Google pro gives you a set rate with a cool down. Then it refills

[brazen frigate]

I am also using qwen and have noticed it has memory loss. Currently working through that issue

[proper turret]

Your understanding is correct.

[proper turret]

Have you enabled the memory hack prompt?

[brazen frigate]

No what’s that

[proper turret]

Prompt:

Enable memory flush before compaction and session memory search in my Clawdbot config. Set compaction.memoryFlush.enabled to true and set memorySearch.experimental.sessionMemory to true with sources including both memory and sessions. Apply the config changes.

[brazen frigate]

What does this do exactly?

What I’ve been working on is it’s
Brain
Heartbeat
Personality
Coding

.

Current Setup:

• Brain: Using ollama/qwen2.5 as the primary model for my thinking.
• Heartbeat: Currently, heartbeats check periodically (every 30 minutes) but can be configured via HEARTBEAT.md.
• Personality/Coding: Configured based on details in SOUL.md.
Speed Improvements:

To improve local LLM speeds, we can tune some settings and ensure the model is efficiently utilized.

1. Use of Local Models: • Continue to prefer using local models for quick lookups and draft work.

2. Resource Allocation:
   Ensure that resources (CPU/GPU) are optimized for running the local models efficiently. This includes: • Monitoring system resource usage (top, htop).
   • Ensuring no other high-resource tasks are running concurrently with critical LLM sessions.

3. Model Configurations: • We can fine-tune model settings if necessary, but typically, default configurations are optimized enough.

4. Preloading Models:
   Preload models in memory (if not already) to reduce initial load times once they're invoked.

[tranquil hazel]

My mac mini is arriving thursday. I'm planning on running openclaw on that via antigravity, using my google pro plan with oauth

i have some gemini tokens to burn

[tranquil hazel]

I have worked with both models in antigravity, but not fully automated

I've ran ralph loops with orchestration tough

gemini is not very good at coding along with a human atm

it will lie and cheat test results

I've read it works much better if you just give it a spec sheet.md

[slim elm]

and why not just get a rpi 5? is the mac mini hype for llama?

[tranquil hazel]

doubt many people are running models on the macs

[proper turret]

Yup. People running a local model.

[tranquil hazel]

wouldn't that be very slow / hard unless you have tons of ram? 😄

[proper turret]

It would still be slow even with tons of ram, since it doesn't have a dedicated gpu

[tranquil hazel]

I got a 24 gb ram 512 storage one for 1020 euro, and that was a good deal

[proper turret]

Case in point, my ryzen 7 8745hs with 780M igpu, with 128gb ddr5 5600mhz ram only has an okay speed for qwen3 8b q4km

[tranquil hazel]

Gonna run gemini models on it via antigravity

[tranquil hazel]

yeh that's not

what ppl are looking for

I don't want to be consuming that kind of electricity 24/7 😄

[proper turret]

Mini pcs, and mac minis are very very low consumption until you add a gpu

[tranquil hazel]

yeh ofc

i have a 3070 in my desktop tower. It weighs more than a laptop

so ofc it'll consume lots

I don't want to run local model, I want to have it work with gemini model

[proper turret]

But yeah, if you're just looking to run openclaw and use cloud api, no need for a powerful machine. I think 2vcpu and 4gb ram would be enough

About $5 monthly if you get a vps

[tranquil hazel]

I also just needed a second computer, and something that is Mac in case I'm making apps for iOS

you need the mac HW for that

[slim elm]

yea irrc oracle gives 24gb ram and 2vcpu for free, thats why i was asking

so that + one of my subscriptions should be good to go

[proper turret]

4vcpu, 24gb ram, 200gb storage

[tranquil hazel]

if the AI skynet apocalypse is coming, I feel I should at least be part of it.

[eternal tendon]

what version of qwen 2.5? coder only replies to me in json..

[echo cypress]

I thought the docs show Oauth, so you can use your subscription, which is subsidized tokens

[tranquil hazel]

anthropic doesn't want you to use oauth for this, and has banned ppl for it.

you could attach it to antigravity

anthropic has a deal with google

you can use claude opus agents in antigravity with oauth

maybe it works maybe it doesn't 😄

[proper turret]

It does work, it's my current setup!

[jade bison]

Right there with you, I've ordered a Mac Mini so I can help support Skynet when it goes down. Doing my part.

[craggy ferry]

Macs all have a “dedicated gpu” in the sense you’re thinking, and they have unified memory.

They run fast enough for the main thread if you’ve got a Studio

But you’re not going to get very far with a Mini, even at 32G that’s not really enough

[proper turret]

I will not argue semantics with you, in my opinion my mini pc has an igpu, 780M.

Mac minis (which is the model being discussed) has an igpu with possibly higher bandwith speeds, still not a dedicated gpu.

[kindred ore]

What’s the method for running stuff, because I was going to buy a crappy server pc with a p100 gpu, and run a model locally, but is there a better way

[vernal river]

This is the way.

[tranquil hazel]

mine is arriving thursday. 1020 euro on amazon.de for 24 gb with 512 storage

bargain

in the meantime

i'll be smoking weed, drinking belgian beers & playing vampire survivors

[tranquil hazel]

no that's a good way. Get a small cheap PC with enough CPU and ram to run google antigravity

then get a google AI plan

[kindred ore]

The free version?

[tranquil hazel]

you can do it free

but for 20 / month you'll get lots

[kindred ore]

K thanks

I’ll try that

[tranquil hazel]

but just set it up free to try it out

[kindred ore]

I will thanks again, do you know how the free quotas are?

[crystal cedar]

Ordered the same config - 24GB bros! 👊

[echo cypress]

I've got an RPI4 with SSD that will be the gateway. I'm currently deciding between running inside docker on the RPI4 or just native, so it can manage the RPI4 for me as well.

[karmic cape]

using a Raspberry Pi CM5 + 2× NVIDIA Spark DGX cluster, and I’m currently testing OSS120 plus four small domain‑specific models with custom ‘intelligent routing’ + embeddings model. Quite happy so far, but want MiniMax M2.1 AWQ to work for at least two users.
It depends on your use case, but if you’re fine with a Linux/Docker setup, it will also run well on a Pi 4 with cloud models.

[tranquil hazel]

I'm gonna call my agent "Henry"

[crystal cedar]

I am deeply honored 🙏

[tranquil hazel]

i still need a screen, mouse and keyboard tbh

[crystal cedar]

there are pretty small portable screens available, perhaps an idea?

[tranquil hazel]

meh better get a cheap 4K screen

[craggy ferry]

I should just jam my agent into a vm on my server cluster instead of depending on one on a Mac, but I want it to be able to look at my iCloud stuff …

I guess the gateway could go on a Linux VM and then the Mac could just run a node?

[kindred ore]

@tranquil hazel do you know how much computer you get for free via antigravity? Also have you tried ai studio

[tranquil hazel]

you get tokens via the google plan

[kindred ore]

K

[echo cypress]

Did you tie the custom routing into the agent loop?

[plain lily]

Which vps would u recommend?

[proper turret]

Actually if you're just running openclaw, just get an oracle free tier vps. All you need is a credit card they can charge $2 and $102 from (instantly returned) for verification, and you get a 4vcpu, 24gb ram, 200gb storage for free forever*

First charge is when you create an account, second one is when you upgrade to PAYG

As long as you're within limits, you will never get charged

[zenith oasis]

I have been toying with a single spark for the last week.

[stone zodiac]

Anyone is running Kimi K2.5 for inference locally ? What is your hardware setup in this case ?

[tiny escarp]

has anyone made a side companion on there desk of a text to speech model or speech to text?

[proper turret]

That would cost... around $400k

[hasty epoch]

2xM3 Ultra 512 GB

for Kimi, can't attest to the speed

[proper turret]

Yeah it would be too slow

[hasty epoch]

guess I'll just kms lol

[amber perch]

im guessing this is the chat for professionals, i need help

can i dm someone who actually knows what there doing and has actually made this work and can explain to me simple questions that i know the answer to but need reasurance

[proper turret]

#users-helping-users

[analog crag]

is it worth buying 2x rtx 3090 for local openclaw setup?

[serene shore]

anyone using local models like glm4.7 or kimi2.5?

[bronze creek]

dont use local models they are not smart enough

even haiku and minimax give bad answers sometimes

[warm slate]

Let's say you run a GLM 4v7 Flash on a 10 year old i5-6500T low tdp CPU with 32G DDR4 at 9k/65k ctx what is the round trip time for a "Hi" telegram message ?

[craggy ferry]

20+ tps too slow?

[zenith oasis]

Just use a model like Claude

[distant mortar]

Trying to set openclaw up locally with ollama what model should i use with my hardware? 7900 xtx 24gb and cpu is ryzen 9 9950x windows with ubuntu honestly could use a few tips setting up as well had it running a couple times but messed up

[proper turret]

any model that is under 24gb pretty much

[craggy ferry]

glm-4.7-flash at like 4 bit might fit

[proper turret]

Dang she hungry

[distant mortar]

is there a certain one that is unlimited since im local with ollama? ill admit im a rookie when it comes to this appreciate the reply

[proper turret]

All of them, since you're using your gpu to run it 😄

[craggy ferry]

yeah i was like what does unlimited mean

[distant mortar]

well i had kimi 2.5 set up and it like stopped working said my ollama was limited?

[craggy ferry]

you definitely weren't running k2.5 with ollama

[proper turret]

You must have been using cloud then

[craggy ferry]

you would, uh, know

[distant mortar]

yes it was cloud

[proper turret]

Gemma3-27b would be best

And i'd recommend switching over to llm studios

[distant mortar]

will also look into that do they work together better on llm studios

[warm slate]

I'm playing with llama.cpp - it "works"

[proper turret]

this is what i use, for vulkan

llm studios and ollama are just more user friendly

[warm slate]

my cluster of 3 thin clients with APUs is the slowest backend possible, but yeah, vulkan + rpc runs OSS 120B with 1-2t/s tg on ddr4 so-dimms

[craggy ferry]

llamacpp works well, give it a lot of memory for context cache, can't use qwen-next-coder tho

at least not the unsloth quants, haven't tried anything else because in a shocking twist i don't actually have the vram to load an unquantized 80b model

[warm slate]

my current setup for the bot is 5 provider/models - each is a llama.cpp instance on a different host - seems to work after the latest update with the "set default" agent models

[flat grove]

What tool is that?

[proper turret]

Antigravity tools on github

[distant mortar]

Hey fam I have another question I am entering my model manually in openclaw onboard is it not google/gemma-3-27b when I proceed to model check it says not found

[proper turret]

Uh, you've loaded it in ollama/llm studios?

as a local model?

[distant mortar]

Llm studio local yes

[proper turret]

you'd need to create a custom provider + list of agents

"models": {
"providers": {
"atbp-proxy": {
"baseUrl": "http://100.127.38.35:29123/v1",
"apiKey": "sk-##################",
"api": "openai-completions",
"models": [

google is so generous, i paid $0 for 1 year for all this

[distant mortar]

so im connected locally and everything with gemma-3-27b-it and gateway is green but i get no responses

[proper turret]

are you using your local endpoint, port, api key, and right model name?

[warm slate]

it might run in timeout if you run the 10k tokens for the first time - any cpu/gpu usage ?

[distant mortar]

◇ Config handling
│ Update values
│
◇ What do you want to set up?
│ Local gateway (this machine)
│
◇ Workspace directory
│ /home/pul/.openclaw/workspace
│
◇ Model/auth provider
│ Skip for now
│
◇ Filter models by provider
│ All providers
│
◇ Default model
│ Enter model manually
│
◇ Default model
│ lmstudio-community/gemma-3-27b-it
│
◇ Model check ─────────────────────────────────────────────────────────────────────────────╮
│ │
│ Model not found: lmstudio-community/gemma-3-27b-it. Update agents.defaults.model or run │
│ /models list. │
│ No auth configured for provider "lmstudio-community". The agent may fail until │
│ credentials are added. │
│ │
├─────────────────────────────────────────────────────

[proper turret]

It's not loaded at all

you need to edit "C:\Users(yourpcusername).openclaw\openclaw.json"

[distant mortar]

\wsl.localhost\Ubuntu\home\pul.openclaw\workspace

like this?

[proper turret]

well not the workspace, go back 1 folder up and you will see openclaw.json

[warm slate]

I spawned 3 new agents - with 3 empty GPT-OSS 20B models - on 3 nodes with same settings, only 1/3 wants to start reading on its own. (?)

[distant mortar]

fixed the directory not sure my model is 100% right keeps making me select amazon-bedrock/google.gemma-3-27b-it

[proper turret]

you can probably find this in llm studio where you loaded the model in

it should say what name to use

and replace amazon-bedrock/ with whatever provider you created

[distant mortar]

lm stuidos tells me this "lmstudio-community/gemma-3-27b-it " is the exaxt model

[proper turret]

sounds good

you need to add a provider named lmstudio-community then

[proper turret]

replace atbp-proxy with ur provider name

[gleaming cove]

I have 2x 3090-24gb vram + 128gb ram, have someone succesfully fit a 70b model on 2x24gb vram + 128ram?

[proper turret]

what quant?

[stone zodiac]

Yeah I figured out it’s impossible so it’s better to use it via api and it’s still la 2-4x cheaper than opus

[bronze creek]

only use local/cheap models if you want a broken system

i wouldn't touch this with less than 120b models, even 480b are going to fail constantly

[gleaming cove]

I was considering about trying 5 or even 6-bit with ram offload but now that I think about it again it sounds not really realistic. 4-bit should be possible on paper - is anyone running that on a similar setup?

[late pawn]

hi, i'm thinking of trying openclaw, got a GMKtec K8 miniPC with 8845HS/780M/96GB with 16GB for igpu + 4TB SSD. Based on what i've been reading i should run a hybrid model with small tasks run locally, and then a paid API? The MiniPC is my main desktop and i would not like it to slow down.. what should i be looking at?

i'm not versed in AI, so much information i'm too old and slow

running fedora 43 on the mini.

so much to learn.. i bet my question is dumb.. also why use the $5 VPS with paid API access to LLM's? if it takes almost no resources why not just run that on desktop?

anyways sorry for dumb questions, i try to read more

[iron sparrow]

I am having issues with signal-cli on raspberry pi 5. Anyone else?

[plain bolt]

so abusing

[late pawn]

guess what i'm asking is, does anyone run local models on amd ryzen 780M + 96GB+ ram? or should i forget it?

i'm confused why ppl are buying mac mini's for this, i assume it's to run everything locally?

yet guides all say use API to LLM models

[proper turret]

8745hs or 8845hs? Do you have a TPU

[late pawn]

8845HS

yes basic TPU

[proper turret]

You can run 8b models comfortably, and 72b models will work BUT very slow

Like nowhere near conversational for big models

[late pawn]

thanks. do you know would the AI in the background constantly use processing power and so spin up the fan? i mean is the best for this system to use hybrid, both local and remote?

[proper turret]

No it will remain super low power usage until prompted

[late pawn]

ok.. i will play around, but for best i guess i should pay for remote 72b model?

[proper turret]

It would be best to use a local model for some simple inquiries, or heartbeat

[proper turret]

Nah, 72b is not as smart as some of the dirt cheap cloud models

[late pawn]

well i don't even know what 72b means :) just thinking best to pay for some cloud model for the heavy lifting

[proper turret]

It's like a general measure of smartness, 72billion parameters it was trained on

[late pawn]

ok.. i'm old and this ai thing is evolving way too fast.. just wanna play with openclaw, see if i can change the way i use desktop

[proper turret]

For local models, always offload the entire thing to gpu or it won't be fun

[late pawn]

thanks.. will slowly test in VM first, then see if i can adapt to desktop.. it would be cool to have

[proper turret]

Use something easy like llm studio to load the local model

[steep wedge]

I am successfully using signal-cli, but on an Ubuntu VM, not a pi.

[acoustic flume]

If not talking about price, why Mac Mini? Are some tools only available on MacOS for the bot?

[blissful kiln]

Thoughts on using a desktop vs. server install of Ubuntu on a VM?

[bitter scroll]

I had to build a couple of dependencies from source but it works great now

[iron sparrow]

ok not just me then. ty @bitter scroll

[karmic cape]

I implemented a custom provider which detect all native domains and Skills which attach flags (experimental). The system automatically identifies ~17 different domains with only 2–4 ms of additional overhead. I’ve worked extensively with fleets of SLMs on edge devices over the past years and am TRYING merging these learnings into the most practical openclaw version, combining local and cloud models or whatever is available.

[karmic cape]

What model/models are you running? vllm?

[potent olive]

Hey ya’ll, I’ve been working on something for a while. No power. No internet.

[echo cypress]

I thought that's what the frontier chat apps were made for... to give you assurance.

[worldly zodiac]

What blackmagicsourcery is this

[dry hull]

https://huggingface.co/TeichAI/GLM-4.7-Flash-Claude-Opus-4.5-High-Reasoning-Distill-GGUF looks interesting

[proper turret]

Using antigravity oauth and proxying it as an openai api that load balances usage

[limber tundra]

hi all, wil lthis run fine on a pi 3?

[eternal pine]

is anyone running clawd against an llm on strix halo or a dgx spark? I'd like to know what kind of performance they're getting with larger context windows

[karmic cape]

I guess the 1GB RAM will be the issue.

[karmic cape]

2x spark, what context windows are you looking at?

[eternal pine]

64k or 128k

its unusable imo on strix, so im really considering buying new hardware. Spark has my interest

or a 64gig mac mini m4 pro

[limber tundra]

Gotcha, Ive been contemplating just getting a mac mini but not sure. is the m1/2/3 chips a necessity or is a i7 gpu fine?

[eternal pine]

kind of the two im bouncing around in my head, but i dont watnt o get another pp bottlencker like Strix

which generally works great for normal queries, but dies a horrible speed death with agentic loops

unless i use a tiny context window, and thats basically useless for anything ubt the most basic tasks

[karmic cape]

On a single Spark, I tested OSS120 about a week ago and achieved ~35 tokens/s with a 64k context window, dropping to ~14 tokens/s with a 40k input. [Runnig currently ](#hardware message)

[eternal pine]

im not super worried about token generation as long as its double digits. How was prompt ingestion speeds at near max context size

for me a 64k buffer on strix with gpt-OSS-120b can take minutes

to first token

thats not a good experience especially if the cache gets invalidated

i've heard the spark has incredible pp speeds, but its hard for me to find relevant users using it for this purpose

and i want ot compare it to an M4 pro at 64 or 128gigs, as thats the same price point basically

well, a lot cheaper up to near the same pp

[tranquil hazel]

to run 24/7

also, fomo

if you get a mac mini, you can send iMessages to your virtual waifu girlfriends.

[eternal pine]

im also confused by it, when most of the people who talk about it are not running inferrence locally. Might as well use an RPI if you're going to use a cloud provider for your llm

[simple rain]

2013 mac pro for the win

[tranquil hazel]

i just got my mini an hour ago lol

glad there wasn't a brick in the box

[eternal pine]

lol

the ol brick indiana jones trick

haven't seen that since the old days of best buy graphics cards

[tranquil hazel]

apple hides their box inside another box

pretty genius

anyway still need a monitor for it so i'm setting it up tomorrow

[eternal pine]

damn it, im really not sure what to do

[olive sleet]

What are the benefits of a Mac mini over a vps?

And do you guys know how capable are the local models that can run on a 16gb Mac mini m4 / m2 pro? To reduce api costs

[eternal pine]

depends on your use case honestly

for Systems Engineering assistance, i wouldn't trust anything smaller than glm 4.7 flash. if you're just doing general life style personal assistant tasks with it, gpt oss 20b will do fine

if you're not using local inference, the only benefit of the mac over the vps woudl be integration with imessage for communicating wiht the bot

[olive sleet]

Thanks 🙏

[tranquil hazel]

I only got it because I wanted to run this, and wanted a second computer, and didn't own a mac system yet. If I want to make iOS apps then I need mac hardware.

I've already made a few things with antigravity the past months. But not for iOS

[olive sleet]

Which one did you get? And are you running any llm locally?

[tranquil hazel]

I literally just got it

not planning to run local LLM

[olive sleet]

Oh alright

[tranquil hazel]

gonna use it attached to antigravity

running models from there with google AI pro plan for starter

[olive sleet]

Is it gonna bill you for api usage for openclaw + google ai pro plan?

[tranquil hazel]

No

That’s why I use antigravity

[olive sleet]

That’s crazy

Do you have a link for a tutorial?

[tranquil hazel]

there's probably tutorials online

hold up

[tranquil hazel]

https://youtu.be/1Jqaj1KN5vA?t=241

google and openAI allow you to use oauth, not api

anthropic doesn't like that

but anthropic has a deal with google

you can use anthropic tokens on google antigravity

[olive sleet]

Tysm 🙏

[tranquil hazel]

google also does some stupid stuff for students with free accounts to get them into the system. Also some cheaper family accounts. So you can set up a system with multiple accounts

using google AI pro (family) plans

I only tried half of it

if I learn more, I'll share here

[fast pond]

Blegh. Deciding if I should get a spark or not to host this locally.

Is there any other cost-effective options? Maybe to save a grand or two?

[eternal pine]

im in the same boat joey

i can tell you as of right now to avoid strix halo

maybe the npu enablement will significantly improve pp, but right now its useless for agentic work with large context

spark looks like a better option, but at the point of spending between 3 and 4k im hard pressed not to just buy a m4 pro or m4 max

honestly just hard to get a clear head to head of functional performance between the two thats not your typical fluff token counting review

write a 500 word story is useless as a comparator

seems like all the ai review slop channels just focus on T/s and writing f**king stories all day.

[fast pond]

I'm thinking I might use my 3090 and 3070ti with a sort of round robin with quite specialized models.

IBM's granite small is pretty awesome with more technical aspects

So I have my subagents run granite right now

[proper turret]

You can do this for any google pro account, you can 6x your limits in antigravity because you'd have 6 pro accounts with different usage buckets.

[tranquil hazel]

yow, tell me more.

I already have google pro acc

[proper turret]

Just google

How to add family members to google one, invite 5 different emails, accept it, then you can use all of them when you run out of usage in 1 account

I currently use sonnet for my support agent

[tranquil hazel]

I'll ask gemini

setting up openclaw tomorrow on a mac mini

probably in docker

[proper turret]

Here you can see my setup

[tranquil hazel]

I wanna go degen with this

I'm used to antigravity

already did kinda stupid things with it

orchestration with parallel ralph loops

but this openclaw thing running 24/7

[proper turret]

That's just an extra tool so you can switch accounts in 1 click on antigravity when you're out of usage. Also you can set up a proxy so you can use it as openai api with load balancing (massive plus)

[tranquil hazel]

sounds really degen, I have to try it

[sweet egret]

Why you use mac mini for open claw?

[tranquil hazel]

the PC I'm using now is a gamer desktop with a 1000 watt psu

mac mini is 15 watt

[sweet egret]

Oh so you wanna save energy

[proper turret]

I have a mini pc, beelink ser8 8745hs for the same reason, 128gb ddr5 5600mhz ram.

I use it to play around with local models

[sweet egret]

I would be happy to have good pc. I need it to compile LLVM

[tranquil hazel]

electricity is expensive in the EU

[sweet egret]

Cheap in Switzerland compared to Germany

[vernal thunder]

what r u guys paying per kwh

me: 29 ct

[sweet egret]

I dont even know lol

[tranquil hazel]

I have photovoltaic

I also don't own a car anymore

just a 45km/h assisted bicycle

[vernal thunder]

no car either. sharing is caring LOL

[tranquil hazel]

I'm the real deal

[vernal thunder]

thats not belgium is it?

[shell nymph]

Are there recommended specs for the gateway? I'm wondering if I can run a docker container on a Synology NAS with very light specs. Most of the heavy lifting should be on the nodes & the model provider anyways right?

[thick python]

hey guys whats the best VPS thats also low cost, i have open claw running on a 1cpu 1gb ram vps, ive tricked it with a page swap, but having next to no ram is not ideal for automations, and i think thats where my problems are coming from.

[thick python]

will my free server get shutdown for more paid users? when i tried to do this my region didnt have any resources already.

[shell kindle]

I was thinking something similar with my NAS

[thick python]

for api, is everyone just pay as you go? or is there some sub i can pay for to get access to more models? i would like to run as cost effective as possible, been using 2.5 flash lite pay as you go, but its kinda dumb for larger tasks, whats everyone using, whats the average cost?

[proper turret]

You need to upgrade to a Pay as you go account, and then there will be resources available. Note that this will do a charge check for around $100, but instantly returned.

Now you can create the always free vps with 4vcpu, 24gb ram, 200gb ssd for free! It is pay as you go, but if you stay within limits (4vcpu, 24gb ram, 200gb) you will never get charged.

[proper turret]

Get google pro subscription, use antigravity oauth

[thick python]

im doing that its just taking forever to actually upgrade. lol.