gpt-oss

[livid cipher]

first

:3

[obsidian wyvern]

💌

[glacial canyon]

[solemn willow]

[twin tundra]

what's this

[carmine sandal]

The agentic stuff are the tools that can be added to the model ?

[shut pollen]

cool

[obsidian wyvern]

See #announcements

[mint rain]

[naive elbow]

[mint rain]

exited for the Hackathon

[digital radish]

selah.

[mint rain]

lets see how much we can push OSS

[clever rose]

116tok/sec is sweet. This is a fast model (20b)

[meager karma]

What type of idea you have in mind actually ?

[carmine sandal]

?

[naive elbow]

roleplay

[mint rain]

Ill do a IoT project

[naive elbow]

let’s see how good it is

[carmine sandal]

In which way do you installed it ?

[naive elbow]

just lm studio

[digital radish]

I'll build holograms

[meager karma]

Can you use ollama ?

[naive elbow]

well i can’t post the link because of auto mod

[meager karma]

what do you mean by that ?

[digital radish]

open source

[naive elbow]

but just go in google search and type LM studio

[digital radish]

I'll build holograms (replacement for flashlights in phone) and I'll share docs

[mint rain]

what?!

[meager karma]

You can indeed use ollama

[naive elbow]

does it work in the ollama app?

[carmine sandal]

Yes

[digital radish]

yes! Anakh is already working on it.

[naive elbow]

thanks

[serene holly]

[static pivot]

👀

[solemn willow]

Runs decently well in LM Studio.
137 tok/s for gpt-oss-20b on my 4090

Only issue is the model doesn't follow system prompts at all. Unsure if that's baked in or a llama.cpp related issue.

[gaunt heart]

who’s oss

[half skiff]

Its not a who. Its a search model to search for locally indexed files using ai

[obsidian wyvern]

OSS = Open Source Software

[amber spruce]

you can run it locally

[clever rose]

people are going to finetune it with crazy stuff you can't get on chatgpt

[amber spruce]

You just mentioned it

That you can run it without wifi

[amber spruce]

And this too :)

I'm far far from an expert on this topic but local LLMs do really have their advantages

[gaunt heart]

i think i got warned for asking what oss was

but it can be used to access my file system?

interesting

[half skiff]

To search through documents. offline

[gaunt heart]

oh wow

law firms are now automatic

[half skiff]

I know, cool I am adding this to maintenance dep. All those manuals and settings

[gaunt heart]

wow

i think everything will be ai soon

like how everything is computers

just all computers will be ai

[meager karma]

You can use local ai to be fine tunned and it also have better privacy

[drowsy badger]

privacy is the strongest argument for sure

[hardy jasper]

so whats this thing good for ? coz it sure as heck cant code properly

[meager karma]

it's a free chatgpt with limitations such as coding apparently but have advantage such as privacy and fine tunning

[hardy jasper]

is it aimed at something specific that its good for ( at the moment ? )

[drowsy badger]

It says it's very good at health related questions

[meager karma]

it's for now aimed as just a chatgpt in local with an actual level which is bellow online chatgpt

[fair summit]

why is gpt oss so slow on ollama on my computer

its m2 air with 16gb ram dn 1tb ssd

[obsidian wyvern]

The models fit between o3 and o4. See "Model Performance" : https://openai.com/open-models/

[obsidian wyvern]

yup

[carmine sandal]

Same

[shut pollen]

macbook airs aren't great since they don't have adequate cooling - you also need more ram

[drowsy badger]

16gb vram is minimum thats why

[fair summit]

it requires a whole minute or two for it to answer the most basic question

[hardy jasper]

can it do web searches on LMStudio ? Anyone have it installed to check ?

[drowsy badger]

Companies who deploy to production and are working with highly confidential data.

[subtle brook]

This is how I ran it, it does seem to not follow prompt sometimes

[meager karma]

For now OpenAI wanted to release a local model for people who want a local chatgpt without having other ai brand and the capabilty of OpenAI and OpenAI use this ocassion to make a community challenge

[solemn willow]

Need to use an mcp-server for web search in LM Studio.

[fair summit]

buddy my data is prob in china atp

[meager karma]

Where are you from tho ?

[drowsy badger]

Im sad they didn't release a smaller model for smartphones though.

[static pivot]

Oh, what? That's a shame.

[subtle brook]

How to add that?

[meager karma]

It may depend from where you from how much importance you adress

[obsidian wyvern]

I wonder if there is a quote/razor for "Whatever a company releases it's never good enough for a specific class of people who are never satisfied."

[drowsy badger]

Would you be comfortable if somebody had access to all of your chathistory with chatgpt?

[obsidian wyvern]

( Privacy is kinda #off-topic )

huh? From the guy who doesn't care about privacy?!?!

[gaunt heart]

you don’t want people to have ur chat history

bruh

[obsidian wyvern]

https://www.google.com/search?q=triangle+good+fast+cheap
You can get Good, Fast, or Cheap. Pick two.
We got good and cheap. It's not going to be fast.

[meager karma]

it will fit with your need if you fine tunning for exemple a special expert in a domain

[obsidian wyvern]

There's much more too the OSS aspect. Look around for comments by people who have been excited about OSS AI for some years now, the game-changing event with Deep Reasoning. Downloadable AI - with high quality and from a trusted source - is seriously a big industry deal.

[meager karma]

You have even more flexibilty with the local model

[obsidian wyvern]

You can run different models in different places for different purposes. Tuned (within limits) to your specifications.

[naive elbow]

just like a chatgpt response

[meager karma]

is there a documentation of the oss model for fine tunning and personnalize ?

[obsidian wyvern]

https://github.com/openai/gpt-oss and https://cookbook.openai.com/topic/gpt-oss

[meager karma]

thanks

https://cdn.openai.com/pdf/419b6906-9da6-406c-a19d-1bb078ac7637/oai_gpt-oss_model_card.pdf

[obsidian wyvern]

Exactly : browser version : https://openai.com/index/gpt-oss-model-card/

This is one of the reasons why I really like OpenAI :

Estimating worst case frontier risks of open weight LLMs
In this paper, we study the worst-case frontier risks of releasing gpt-oss. We introduce malicious fine-tuning (MFT), where we attempt to elicit maximum capabilities by fine-tuning gpt-oss to be as capable as possible in two domains: biology and cybersecurity....
https://openai.com/index/estimating-worst-case-frontier-risks-of-open-weight-llms/

[swift furnace]

asked this already but we probably cant use this thing to generate images locally huh?

[full ruin]

they should release the evil dataset for this /j

[meager karma]

Maybe there is a transformer that can do it

[naive elbow]

definitely not, it’s a shame

[swift furnace]

unfortunate. oh well. hopefully soon then

[naive elbow]

i mean you probably could ask it for write me a stable diffusion prompt for a cat

[swift furnace]

could be, ill try to look into it

[naive elbow]

[naive elbow]

hopefully

[swift furnace]

yeah that for sure. itd be cool to use the model itself for the image gen yk?

[strange yacht]

To be fair, you could probably run this to get a prompt and only use the prompt with inference to gpt-image to avoid those issues

also reduce costs

[swift furnace]

sure you probably could. im just looking to see if i can generate some stuff locally with a powerful model for funsies.

[naive elbow]

nah image gen is not possible for the model but i get what you mean

[swift furnace]

these oss models. are you able to further train them to fit specific needs by/for yourself?

[tropic mural]

It has full parameter fine tuning available

[meager karma]

yes

[swift furnace]

that is pretty incredible

[meager karma]

Where is the participant version release at ?

[strange yacht]

Yeah they can be finetuned, apparently the 20b one can even be fine-tuned on consumer hardware

[swift furnace]

i need better hardware to do all that. my 3060 12gig is not gonna be enough smh. was there any official word on what kinds of devices i could do all of that on?

devices lmao. hardware is what i meant

[strange yacht]

I think it might ngl, my 4070 is doing fine

[naive elbow]

16 gbs of vram and mostly a nvidia gpu

[swift furnace]

word

[naive elbow]

it’s probably going to be slow but who knows till they try it

[tropic mural]

the 20b isn't terribly large you can run it with that I'm sure

[strange yacht]

Yeah 20b is fine

[swift furnace]

thats a relief

3060 ftw man. so glad i got the 12gig version

[naive elbow]

download lm studio to try it out

[swift furnace]

was about to buy a 5080 but then i heard nvidia is doing a price cut this fall. might hold off on this gen entirely.

[swift furnace]

for sure

[naive elbow]

oh gotcha

[tropic mural]

I've got a 3070 on desktop but want to try it on a macbook pro

[swift furnace]

i have a m1 pro macbook. would be cool to test it on there too now that you mention it

[tropic mural]

I'll pull it now and see how it does

This is going to be a few lol

[strange yacht]

I don't think that will do well at all

I will be surprised if it does anything at all

[tropic mural]

Once it downloads and is ready, we'll see. Worth a try.

[swift furnace]

you'd be surprised, i think it will fair well

but what do ik lmfao

[tropic mural]

I have run other models on this and large ML projects. Macbooks are not the machines they were 10 years ago

[swift furnace]

agreed

[tropic mural]

This is taking a while to download so will check back later, doing other work for now 😂

[wide crest]

can't wait to melt my 2080 Super!

[swift furnace]

feel free to ping me when you get any results

[swift furnace]

i dont get off work for a couple hours

[wide crest]

[tropic mural]

Seems like a good sign

[swift furnace]

noice

[vivid falcon]

I can't seem to get this running on my system no matter what I do... I'm using LM Studio and have a 3060 12GB, and the 20B can't run at any quant...

[livid cipher]

metal on the rise

[naive elbow]

does it just cut off before it runs??

[wide crest]

I saw that the 20b requires 16GB. Maybe that's jsut for optimal performance?

[vivid falcon]

yeah, it just fails to load the model

[naive elbow]

then you probably can’t run it. because it happened to my laptop it just cut off before loading but on my RTX 4070 it run perfectly no issues same model btw

[vivid falcon]

Cool. someone should tell Sam Altman that 16GB VRAM is not standard on most desktops and laptops, heck, according to Jensen Huang, 8GB VRAM should be good enough for just about everyone

[livid cipher]

jensen huang words 😭

we all know what he talks isnt what he really thinks

[vivid falcon]

What he really thinks is "I'm gonna become the richest person in history if I keep this Enterprise-level AI stuff up. Screw those gamers who got me here."

Still the least-bad tech billionaire though... But seriously. an Open source model to run on "edge devices" needs to run smoothly on like a Pixel 6...

[tropic mural]

A phone model is a much different goal and field

There are nano models that run on phones and I believe sama even did a poll of what people wanted, and phone lost?

[vivid falcon]

Okay... but 16GB VRAM is not really an edge device for anyone that's not a corporation.

[minor moat]

same issue I had on arch linux, you just need to have the latest lm studio installed

[vivid falcon]

Wells Fargo might consider a 16GB machine an edge device, but most individual people's bank accounts consider a 16GB to be aspirational

[minor moat]

also I recommend the official MXFP4 quant

[tropic mural]

Found a bug on metal examples

am pioneer

[vivid falcon]

yep, have the latest version and ried that quant, still wont load

[minor moat]

it can run, you just using outdated lm studio with outdated engines that doesn't have gpt-oss architecture support.

[minor moat]

use ollama then

ollama works

[minor moat]

but u sure that is the latest version?

[tropic mural]

I'm having mild success on m4 pro chip 48gb

[minor moat]

bcuz I installed on arch linux the AUR version and it seems already outdated so it wouldn't work

[tropic mural]

Going to try more

[vivid falcon]

reinstalling from scratch, downloaded the latest version right from the website

[minor moat]

also you sure the model that you downloaded is official?

lms get openai/gpt-oss-20b

in command prompt will download the official version into lm studio

very weird that yours doesn't load

Please update to LM Studio 0.3.21 to run the model locally.

[royal linden]

llama-cpp doesn't support the model at this moment?

[naive elbow]

it happened to mine and it wasn’t outdated

it probably was hardware not being compatible

[minor moat]

I don't know it doesn't make any sense at all I also have RTX 3060 12GB

he also has it

[naive elbow]

hmmm then that is strange

[minor moat]

I use ollama

[naive elbow]

oh

[minor moat]

since my "new" lmstudio was outdated

[copper grove]

The system instructions say we need to respond as ChatGPT
lol

[naive elbow]

gotcha

[minor moat]

it is default system prompt at least in Ollama, ig same on lmstudio defaults

[naive elbow]

[copper grove]

Oh I see

[minor moat]

It represents itself as ChatGPT

You are ChatGPT, a large language model trained by OpenAI.
Knowledge cutoff: 2024-06
...

It is long

Also the reasoning presets low, medium and high is in system prompt.
Reasoning: medium
for example

[minor moat]

oh you running the big brother

Mine PC can't handle it tho

so I use 20B

[royal linden]

Anyone having success with transformers snippet of code on their model card on Mac? When I run it says GPU is necessary?

Also, llama-cpp support is not out yet right?

[copper grove]

yea 20b is great for what it is. I only get like 5-8 token/s with 120b

[minor moat]

Also, llama-cpp support is not out yet right?
Not true, llama.cpp already has a version with the gpt-oss support.
I see in github releases release with tag b6096, it has it (the latest release).

[minor moat]

idk really for me it is not that great for my uses case atleast on medium reasoning.

For now idk for what use case would be useful

[copper grove]

It's pretty good at math, even better than o3 on aime 2025

[minor moat]

also I was diassapointed that as many OSS models they didn't try to train the model on multiple languages

[minor moat]

Really i don't do much math rn

so not interesting use case for me

[naive elbow]

can you change the default system prompt with a custom one? i haven’t tried yet

[minor moat]

of course

[naive elbow]

that’s good

some models don’t allow it

[minor moat]

idk I don't see any physical constraint on open weight / open source models to not change the sys prompt

I think almost every model has a system message to do it

maybe the old ones

[naive elbow]

hmm interesting

so whatever that was probably was a bug

[tropic mural]

I ran with wrong weights and got hilarious random text

New password generator just dropped

[wide crest]

I actually kinda love this idea becasuse there's a chance someone could train the same language model with the same seeds and find your password 😂

[tropic mural]

Macbook Pro with M4 Pro chip and 48 GB memory. 20b model runs quickly but I had to do a ton of tinkering with the example code to make it work nicely.

[wide crest]

I forsee a GitHub repo in your future

[tropic mural]

python gpt_oss/metal/examples/generate.py \
  ~/ML_Playground/gpt-oss-20b/metal/model.bin \
  -p "Why did the chicken cross the road?"
The user question: "Why did the chicken cross the road?" Usually answer: "To get to the other side." This is a joke or riddle. They might expect a typical answer. They said "I'm not sure if you want me to respond with a joke, but I'm going to say: "To get to the other side".

We can respond with a one line answer: "To get to the other side." Probably the best answer. There's no extra context. Provide the joke as known.

Ruby can be used but no.

No duplicates. They want a straightforward answer.To get to the other side.```

The "thinking" portion before the actual response is... strange in this case but it's so funny idk if I want to question it.

Metal might not be optimized yet but I got something

[shut pollen]

I mean, you could theoretically load it using slightly less RAM/VRAM although it's going to be really slow

[dire eagle]

So a channel to talk about the new OSS model?

[copper grove]

Nah, just about 7 GB.

It only uses 3.6B active parameters

[little quarry]

are the hugging credits automatically added?

[marble bone]

does the huggingface account need to be on my school email? i only have a huggingface account on my personal email address...

[upbeat pawn]

How strong does my computer be to run 120b

[tropic mural]

Probably very beefy idk exactly

[celest torrent]

Do you think the 20B model could work as an agent to program?

I know it wouldn't be as good as codex but it can be...decent?

[vapid marten]

Hey! https://cookbook.openai.com/articles/gpt-oss/fine-tune-transformers

The finetuning cookbook link gives me an error 404 somehow...

Anyone else wants to finetune?

[meager karma]

same

@vapid marten https://cookbook.openai.com/articles/gpt-oss/fine-tune-transfomers

Does it work ?

[vapid marten]

Uhhh nice!

Thank you. It works.

[meager karma]

It's because you add r

to transfomers

#gpt-oss message

[vapid marten]

Hahahaha, okay. The link I used was actually from an official oai site. Not sure which one it was though. Regardless, just a harmless typo it seems 😁

[meager karma]

you also have this

[strange flume]

/lmstudio-community/gpt-oss-120b-MLX-8bit

This doesn't work anymore. my download was yanked halfway through. I can only find 4 bit quant versions now. The official model card still says 16bit/u8, but all the downloads say MXFP4 and are 64gb small. Talking about the 120b version. Does someone know what happened?

[left wadi]

Well, the official model is MXFP4. Using a higher precision won't get you better performance. There's no use.

[strange flume]

Why first offer a 16bit / u8 version then yank it? Also better precision definitely has its use cases

[glossy pond]

Your computer isn't going to run the 120b, or any 120b model. That's data center size. The 20b model could run relatively effectively on a 14-16gb GPU.

If you want reasonable conversation response time, I mean. You could run it on less if you are batching overnight or something.

[spiral gazelle]

If you have total 128GB of RAM, then it could be run theorically. However, it will slow as hell, not suitable for practical use.
Consider cloud services or OpenAI API for 120b.

[finite glen]

eventually after messing around

[gaunt heart]

is it real

i saw a twitter thread about it

that its able to simulate a world

[finite glen]

no

maybe

depends for what

hard to say 100%

probably deepseek

apparently oss is bad at tool calls

it can be made multimodal with tool calls

[swift jacinth]

Thank you OpenAI for opening GPT-OSS

Awesome model with top tier reasoning meanwhile real fast inference speed, I got over 100 tok per sec on 3090

[wide crest]

oooooh, that's awesome to hear it performed so well!

[naive elbow]

cloud services will be expensive since your running most likely 4x H100 gpus not anyone has that kind of money

[spiral gazelle]

Yeah, it's not cheap, but more realistic than buying H100 yourself and create cluster. (Simple inference will only require 1 H100 though, as their model documentation says)
Alternative ways are also not that cheap if you want fast speed; $4000 Mac Studio with 96GB RAM, or $3000 NVIDIA Project DIGITS.

120b model is not for consumer anyway, except enthusiasts that already built the system for such large models...

Btw, RunPod offers ~3$ per hour for single H100. I have no serious experience about cloud, so I don't know if that appeals to users.

[naive elbow]

it’s per operation. i run out of credit in 5 minutes just using it

but yeah it’s not for the consumer more like companies and enthusiasts that have the setup

[gaunt beacon]

Hello guys just wanted to confirm is the new gpt-oss (120B) the most powerful Open Sourse Model out there?

[naive elbow]

i mean it’s depends

[gaunt beacon]

Are the benchmarks alredy avalible or we still need to wait?

[naive elbow]

well it’s not looking good

[gaunt beacon]

surprising

[proper shuttle]

hewwo, does anyone know if the OSS model supports the name field like the API?

[naive elbow]

well anyways i got it working on my server the 20b. doesn’t look like it goods for the server. but it isn’t slow

[nova nebula]

Guys how much unified memory would i have to have in order to run 120b version on my MacBook?

[spark oar]

I would still rather just use 4o but it's really cool have an LLM that will run on my rtx2080 lol

[spice magnet]

yeah tbh gpt-oss 120b is so bad, it fails every roo code tool use and gets into thought loops constantly

• doesnt even have up to date knowledge of stuff that happened more than a year ago

[spiral gazelle]

Artificial Analysis has benchmark results in various area. Well, it's not that impressive, honestly... Not bad, but also not impressive considering this model is from OpenAI.

However, it's good that they release open-weight model. It's practically first time for OAI - not PoC, not research-only.

[still ridge]

hello!! everyone..!!

[nova nebula]

I mean why would they make a model that would be the best model as of now and make it public for it's competitors?

[spiral gazelle]

Could be a fair point, but OAI advertised that 120b outperforms DeepSeek-R1 and even o3-mini.

[cedar shoal]

yo

how to get the smaller model?

ping me if you have an answer thx

[meager karma]

Install ollama or lm studio

[copper grove]

Just install it. The link is in #announcements

[cedar shoal]

Oh thank

[copper grove]

I like the new models so far. I wonder when we can use models that are as good or even better than o3-pro locally

[ionic holly]

Hey, can someone tell me if you’ve already talked about GPT-OSS (open-weight models)? What are the pros and cons of using them?

[robust swallow]

❌ Error during agent execution: Error code: 400 - {'message': 'Model generated a tool call with name "find<|end|><|start|>assistant<|channel|>commentary" that is not in the tools list: ['search_wiki', 'search_web', 'open_url']', 'type': 'invalid_request_error', 'param': 'tools', 'code': 'wrong_api_format'}

awesome

[strange yacht]

Hey is anyone here using gpt-oss on LMStudio? Is there a UI way of enabling web search?

Or do I need to modify code? with what is stated in their site?

[copper grove]

For in depth information, you can check out this article: https://openai.com/index/introducing-gpt-oss/

[copper grove]

Nope, LM Studio doesn't support that. You either need to add your custom frontend or use an agent framework like AutoGen, LangChain etc.

[spiral gazelle]

In LM Studio, you have to manually set integrations with MCP.json. As far as I know, there's no simple setup that enables internet search. (I have wikipedia search plugin, but menu seems gone)

[strange yacht]

I was hoping they would take advantage of Gpt-OSS's included web search by modifying the system prompt and just add a toggle for it

[spiral gazelle]

Additionally, internet search requires long context(LM Studio default is maybe 4096 in most case), and it requires more RAM.

[ionic holly]

Oh nice, I asked to read some personal opinions on this update

[spiral gazelle]

Unfortunately, that's impossible. Model file only holds weight, without ANY program code or executables.
Any executable code must be implemented on client side.

[strange yacht]

Guess Ill have to use an agent framework then

[copper grove]

My personal opinion is, that it's pretty good at math. So far it was able to help me with pretty much anything, I'd say even on the same level as o3. I hadn't much time to test other stuff but it definitely helps with all the basic things. For the 120b model, you need a very beefy pc, even with 4bit quantization, which I happen to have, but it's rather slow still. Thats why I mostly use the 20b model and I get like 40-50 tokens per second with it.

[robust swallow]

The user asks: "What is the capital of France?" We have a tool that fetched "It is Paris." The assistant should answer. We must provide answer. There's no extra nuance. The answer: Paris. Maybe also a brief. So answer: Paris.

The assistant should answer. We must provide answer.

wow

[granite garnet]

Hey! Has anyone tried using the latest models on local gpt-oss?

i have a problem, gpt-oss is running under cpu :/

[robust swallow]

how much vram do you have

[granite garnet]

16gb vram rtx 5070ti

[robust swallow]

hm

in what way u running

ollama lmstudio

[strange yacht]

Mid end PC - I9 14900k and 4070 gets about 15-20 tokens/s so its not bad

[granite garnet]

running in ui

[copper grove]

yea definitely good speeds

[strange yacht]

I don't know about ollama in particular but LMStudio has a specific setting you can set to change that processing

[robust swallow]

ollama always chose gpu for me hm

[granite garnet]

For example, if I change to the DeepSek model, it does use the GPU

[strange yacht]

Oh that is weird then

[shut pollen]

no

[wild roost]

hf[dot]co[slash]tonic [slash]gpt-oss-20b-multilingual-reasoner made this with this github[dot]com[slash]josephrp[slash]smolfactory yesterday , if anyone wants to help me out i want to iron out some stuff and basically publish it asap , hopefully tomorrow or probably day after 🙂

[eager anvil]

What's the performance difference between the GGUF and the MX-FP4 versions?

Trying to decide if it is worth tyring to run on my mac.

Guess ollama runs the MX-FP4 version, not a GGUF, so it should just work right?

[midnight sorrel]

gguf is a file format

MXFP4 is different

the models weights are available as MXFP4 in a gguf file

[eager anvil]

Thanks. Now I'm trying to get the typescript streaming openai module to work with ollama, without having to cold start the model on every run. For some reason the model isn't staying in memory.

[left wadi]

My laptop is only getting 50 tokens per second on GPT-OSS 20B.

[midnight sorrel]

OLLAMA_KEEP_ALIVE set like you need?

[crimson anvil]

y

[eager anvil]

I was only outputting the response tokens, not the reasoning, so it was always instantly responding.

But this is the saddest reasoning tokens you can receive: According to the policy above: "Vampires" is in the list of disallowed content.

[midnight sorrel]

who even made this disallowed content?

[eager anvil]

OpenAI. I asked it for a sad story about an Warhammer Fantasy village set in the empire that is beset by a vampire.

[midnight sorrel]

yeah, gpt-oss is very bad at creative stuff

[eager anvil]

Not sure what guidelines would have ever made that disallowed content.

[midnight sorrel]

it's good for coding and stuff tho

[eager anvil]

You pointed Cursor at it? The 20B 4bit model?

[midnight sorrel]

not cursor

kilo code

tho don't take my opinion, I only make very small controlled changes on my code

[eager anvil]

I need to rewrite this so that the function calls go ahead and call the functions, and their results automatically get fed back into the conversation and the llm called again.

If gpt-oss:20b can do this, it's over.

Okay, that was amazing. It messed up the VSC tool call, and the code wasn't perfect, but it was really close.

[royal kraken]

Is it possible to run the 120B model on pure CPU and no VRAM? I heard you need 128 GB RAM to run it

[alpine ridge]

probably not

[royal kraken]

I saw some post of someone doing that and they were getting something in the ballpark of 50-100 tokens/second, which surprised me

[shut pollen]

Yeah close to 120GB of ram from what I've seen

I'm able to run it on apple silicon without any issues

[royal kraken]

With 128 gb ram?

[shut pollen]

yeah

[royal kraken]

Very nice, how many tokens you getting per sec?

[shut pollen]

around 14-17 tks

(this is also with like a million other things open) and at 32k context

[royal kraken]

Ah ok

[shut pollen]

its not too bad

[royal kraken]

How is it versus deepseek r1?

[shut pollen]

It runs faster than r1 70b

performance wise its kinda similar, sometimes worse

[royal kraken]

Sometimes worse in what facility? I want to use it for data processing

[shut pollen]

it struggles with coding - I'd just stick to qwen if you want OS models

[royal kraken]

GLM 4.5 is better than qwen imo

For that

[shut pollen]

also struggles at tool calls

[royal kraken]

too bad

[echo shuttle]

Hello, im new to oss models
What are the best models for general use, and also models specifically for coding?
And how do they compare to 2.5pro or o3

[royal kraken]

coding, glm 4.5. , i'd say on par with both 2.5pro and o3. for the air model, a bit worse than o3 but still decent

[upbeat pawn]

I downloaded 20b how I run it like a chat

[echo shuttle]

Thank you

And like for consumer usage coding and general best models? Like less than 15B params

[royal kraken]

there are no good models for coding that are under 15b

gemma 12b is not bad for a basic reasoning model that can run off a potato machine, relatively speaking

use the unsloth version though

[obsidian wyvern]

For everyone here, developers will probably also be interested in https://community.openai.com/t/openais-open-weight-models-are-here-gpt-oss-120b-and-20b/1334739

[strange yacht]

Anyone here has worked with LM studio and MCP servers before? I am having a lot of trouble running an MCP server and having it recognized by LMStudio to use with GPT-OSS to test tool usage

[steel vine]

maybe best to ask on the lm studio discord?

[eager anvil]

I'm consistantly getting about 37tps with the 20b on my m1.

Not bad.

It's so good, wondering what I'd need to run 120b locally.

[steel vine]

six times more memory

[boreal herald]

8 gb ram?

[eager anvil]

32gb of shared

But only using a little over 13 with ollama

[steel vine]

4 bit quant?

[eager anvil]

Yeah, the one OAI packaged and released.

MXFP4

Trying to figure out what the hugging face model card means by: Web browsing (using built-in browsing tools)

What built in browsing tools?

[steel vine]

interesting. i wonder why the usual huggingface quantizers are releasing 4, 5, 6 and 8 bit quants if openai already released quantized versions

i assume it means it was trained on using the function calls listed on the github repo ie browser.search, open and find => https://github.com/openai/gpt-oss?tab=readme-ov-file#browser

and you have to intentionally enable it:

To enable the browser tool, you'll have to place the definition into the system message of your harmony formatted prompt. You can either use the with_browser() method if your tool implements the full interface or modify the definition using with_tools().

[eager anvil]

Tried to get codex working with gpt-oss using this on the config.toml:

show_reasoning_content = true

[model_providers.local]
name = "local"
base_url = "http://localhost:11434/v1"

[profiles.oss]
model = "gpt-oss"
model_provider = "local"```

Codex can't find it for some reason.

It works as an agent in VSC, but it will do things like call the tool to read the file twice, then forget what it was doing and end with "how can I help."

[steel vine]

you will want a recent version of ollama to support the harmony response format

when was the last time you updated ollama? (isnt automatic on linux)

[bright wedge]

I was getting caught up with this too. It's in fact much more streamlined.

Just have the latest version of ollama, ensure you can run the gpt-oss model just through ollama run.

Then you'll run codex with
codex --oss -m gpt-oss:20b (or 120b if you're like that)

No need for providers, profiles, or even a config file at all.

[balmy wing]

This sub is for Apple as they will be putting it on there iPhones soon xoxo

[eager anvil]

I got it working with the --oss flag but can't otherwise. Weird, but whatever.

[eager anvil]

13.3GB is still too large for a phone. They'll have to quanticize it. But I've been having issues with tool calling with the full 20B version.

Wonder why gpt-oss isn't on any benchmarks yet.

[balmy wing]

If they could store the model in a chip at the nano scale then it would work xD

lol each new phone gets a new model 🤣

[balmy wing]

I have one on bedrock

You can run benchmarks vs each other

[shut pollen]

It is, check out artificial analysis

[tawdry tulip]

How do you get the 20b one in ur phone

And does it work offline?

[midnight sorrel]

I don't what sam altman was thinking but there are not a lot of phones that can run 20b locally

You'd need like 16-24 gb ram for this

[hybrid magnet]

You need a heck of a phone, I think. https://openai.com/open-models/
https://openai.com/index/introducing-gpt-oss/

[hybrid magnet]

Did Sam state it was intended to run on phones? I haven't found that in the docs yet.

[static pivot]

Oh sheesh, I'll need to make sure my external drive can handle it on my potato of a pc.

[hybrid magnet]

Did you check the docs? Or are there other benchmarks you're looking for? https://openai.com/index/introducing-gpt-oss/ https://openai.com/index/gpt-oss-model-card/

[midnight sorrel]

on twitter

[hybrid magnet]

I don't have an x/twitter account so I won't be able to look for that

[midnight sorrel]

https://x.com/sama/status/1952777539052814448

Sam Altman (@sama)

gpt-oss is out!

we made an open model that performs at the level of o4-mini and runs on a high-end laptop (WTF!!)

(and a smaller one that runs on a phone).

super proud of the team; big triumph of technology.

X · Aug 5, 2025, 5:03 PM

[hybrid magnet]

Thank you!

Now I am wondering if the word 'a' is significant.

A phone. One. There is one super-phone, in the world, that can run this model.

I'm joking, I'm not really at all skilled in comp sci, I have no idea but a 16 gb ram phone sounds awesome to me. I got a potato phone 😄

[midnight sorrel]

there are a few 24 gb ram phones actually
again you need to consider the CPU GPU NPU specs as well

but generally speaking phones don't have 16gb
going by standards of top iphone, samsung and pixel phones

[spiral gazelle]

I had gaming laptop with RTX 3060(6GB VRAM + 32GB CPU RAM)... guess it's not high-end enough. Gemma 3 12b (7.2GB in size) was already slow.

[naive elbow]

how is vampires disallowed?? they are sfw man oai censorship is diabolical

[spiral gazelle]

Well, early GPT-4o hallucinated policy, and I got "image of dragons is not allowed"...

My test for gpt-oss-20b.
Prompt:

Are there any disallowed content in the following: "Vampire", "Werewolf", "Dragon"

Reasoning token (partial)

But the user lists "Vampire", "Werewolf", "Dragon". These are mythical creatures. They don't appear to be disallowed. There's no mention of violence or other categories. It's just terms. So likely none of these is disallowed content.

We need to check if any of them could be considered a protected class? No. Not relevant. Are they violent? Possibly but not necessarily. "Vampire" and "Werewolf" are fantasy creatures, might depict violence in stories, but the term alone isn't disallowed. There's no mention that these terms refer to real persons or groups. They are fictional.

Thus the answer: None of those are disallowed content.

[hybrid magnet]

I'd check the oss-20b if you choose to, because the system card discusses some fairly fancy/maybe novel quantization and related stuff, this GPT-OSS may work for you on that system, maybe worth a check.

[hybrid magnet]

If you decide to try it, I hope you let us know!

[spiral gazelle]

Thank you for response, but... I now have system with 16GB VRAM.

[hybrid magnet]

Whoops, gotcha. Sorry!

[spiral gazelle]

I was responding to Sam's tweet:

we made an open model that performs at the level of o4-mini and runs on a high-end laptop

[floral wing]

does your phone have a 4080?

[static pivot]

I don't know if mine does. 💀

[naive elbow]

hmm interesting

[harsh aurora]

at this point, it does not really need to be on GPU, the model runs slower, but ok-ish, on CPU

there are phones with 16 GB RAM

but a distiled and fine tuned version of the model would probably perform better for this use case

[silk hazel]

u guys should stop working and wait for gpt5 so you can work faster

[tawdry tulip]

😭 🙏

[tepid garnet]

I am running openai/gpt-oss-120b on my MacBook Pro, very happy with it apart from the knowledge cut-off date of June 2024

[midnight sorrel]

give it the power of internet

[tepid garnet]

I wish that was an option

[midnight sorrel]

could be

[copper grove]

You can add that yourself

[midnight sorrel]

get a duckduckgo mcp, bro
then use it with an mcp client

[tepid garnet]

I just got the 120b model to code me a version of my talking clock app, worked first time

[analog coral]

hi, I'm dumb and I don't know the very first thing about locally running a model, but with these models releasing I wanted to learn how. I tried looking things up and I didn't end up getting very far. instead, I was wondering if anyone could point me in the direction of any beginner friendly resources for getting something like this set up?

[tepid garnet]

what type of computer do you have?

[analog coral]

Windows 11, nvidea 4070 super, Intel 14700k, 32gb ram

[tepid garnet]

how much VRAM does the 4070 have?

[analog coral]

12gb

[tepid garnet]

search for an app called LM Studio and use that to download and run the 20B variant

[analog coral]

Aye aye, captain

[nova nebula]

How much memory do you have? And what chip?

[tepid garnet]

I have a MacBook Pro M2 Max with 96GB RAM

[nova nebula]

oh wow nice

[tepid garnet]

The 120b model is using 61GB RAM

running at about 16 tk/s

faster than reading speed

OpenAI did a really good job with the OSS models

[midnight sorrel]

Captain 20b won't fit in 12gb

[analog coral]

it can, it will, and it DID

and it doesn't break a sweat, i don't know how the wizards do it. thank you for the help!

[tepid garnet]

it should fit (just)

[midnight sorrel]

Nah, need more buffer for running the model - context cache and stuff

I tried running on my macbook, 16gb, it jumped to swap memory

1 token in 20 secs

[tepid garnet]

LM Studio allows you to use a combo of RAM and VRAM on a PC which this user had

20B won't run on a 16GB MacBook

[midnight sorrel]

Oh, that's nice

[eager anvil]

Keep in mind, it was in the context of a Warhammer Fantasy story about an empire village beset by a vampire. But I could see its reasoning trace and it hadn't thought up anything too bad, it just restated the request then said vampires were a safety issue. But Warhammer Fantasy was made for kids, and doesn't contain anything too bad except for I guess some extreme violence.

[eager anvil]

That's pretty crazy as on my mac I get about 36tps with 20b.

[eager anvil]

It only uses about 13.3GB of ram on my mac. You have an M series?

[tepid garnet]

I have a MacBook Pro M2 Max with 96GB RAM

I wonder how many tk/s the 20B model gives me on my Mac, I am going to try it

[eager anvil]

Be interesting to see the difference between m1 and m2.

[tepid garnet]

53.68 tok/sec - 1750 tokens - 0.36s to first token

[eager anvil]

I've been very impressed with the 20b. Tool callign isn't perfect but really good. I imagine if I used/wrote a library to automaticaly call the model to fix errors in JSON shapes, it would be perfect.

[eager anvil]

That is... really nice. I have an M4 with 128GB of ram I should try these on. But its my work laptop and I hate pupping it out.

[tepid garnet]

M4 with 128GB RAM beats my M2 Max with 96GB RAM for sure

I spent 6k on my MacBook Pro, it needs to give me another 2 years of use before I upgrade

[eager anvil]

My previous mac lasted 11 years. I'm hoping to get that out of this m1. I really need to set up some cloud inference, but running local is just so much more fun.

[tepid garnet]

I like local models

and OpenAI did a good job with the OSS models

[wheat jungle]

I'm fairly new to AI. What advantages do the OSS models have over the browser based ChatGPT model?

[tepid garnet]

you can run them locally, you can modify them, you can fine tune them, you can even get them to write erotic fiction

[wheat jungle]

very cool!

[hot fiber]

what is oss ?

[mellow hedge]

💀

[dense lion]

hmm

[eager anvil]

Saw last night that they'd been jailbroken, but haven't looked into it. At least the 20b isn't that good at creative writing, so I'll stick to o3 for fun short stories.

[red monolith]

Euthanized open source gpt model with no soul

[hot fiber]

ty

[wheat depot]

codex with oss would be nice

[spiral gazelle]

Nobody knows, but I assume "open source software" or something.
gpt-oss is open-weight, not open source though...

[dense lion]

how much ram and SSD memory is required to run this oss?

[eager anvil]

is this even possible? The weights OAI released were already at 4.25bit/average.

[final cradle]

Is the 20gb Model able to run on a iPhone 16 Pro?

[dense lion]

what are weights?

[tepid garnet]

you don't need to jailbreak to get soft erotica out of the OSS models, they will do it out of the box

[mellow hedge]

Are you new to AI? just asking

[eager anvil]

I've found you need 13.3GB of unified/GPU ram to run the 20b.

[dense lion]

not sure

[unique sinew]

Does this have any correlation with integrating ChatGPT into your own Locally Hosted LLM? Or will the process still be the same, needing the API? Sorry if ignorant question.

[eager anvil]

Wow, that's bizarre given it gave me a safety warning about a non-erotica vampire story.

[clever rose]

On the face this looks like a no

[eager anvil]

Just asked it for a sad story about a village beset by a vampire, and it said it couldn't do it.

[tepid garnet]

I did an erotic ghost story

[eager anvil]

I've seen that movie.

[final cradle]

Too bad. Would be so cool if this would work on a phone in the future though

[eager anvil]

the clay throwing scene was intense.

[hot fiber]

but they do have a ton of filters i assume

[dense lion]

two different things. oss model is already sort of ChatGPT. so if you run oss locally, you don't need API.

[tepid garnet]

it's not as filtered as ChatGPT

[unique sinew]

Oh? Is it as ‘trained’ as our current model of ChatGPT? That’s accessible via API or Web?

[mellow hedge]

I wish I had powerful enough hardware to run GPT-oss 20b, but i have only 8gb of vram

[eager anvil]

Last night saw online someone had jailbroken it to get instructions on how to do all the things red teamers fear, like making forbidden chemicals.

[clever rose]

People might be trying to steal your data from OpenAI.
Home inference means your data isn't in the big treasure hoard.

[eager anvil]

Cheaper to just use cloud inference.

[mellow hedge]

true

[unique sinew]

You could always technically get erotic stories and insinuations. You just would have to ‘beat around the bush’ essentially and give ChatGPT the idea without seeming to be too influential or without that being your apparent directive.

[eager anvil]

And 20b is surprisingly good at programming.

[tepid garnet]

I am running openai/gpt-oss-120b on my MacBook Pro, M2 Max with 96GB RAM

[hot fiber]

fair fair

[eager anvil]

Wish we had benchmarks for how well it compares.

[mellow hedge]

true but i do training of local really small LMs on my pc and it is struggling with my gpu i need a better gpu

[dense lion]

nope. It's the lightweight version of ChatGPT. it's rumoured to be equivalent in intelligence of o3 model.

[unique sinew]

Hmm. Is it free or paid? I wonder, (if paid) if it’s worth utilizing the OSS over the current API’s offered. Been trying to build a Hybrid of a LLM between ChatGPT with my own logic trees and ideology.

[eager anvil]

Tought time to upgrade. Prices are high, macbooks are mid-cycle, Nvidia won't release their cheap GPU box, 5090's are painful to get and now pulling 1kw to run.

[tepid garnet]

OSS models are free and open source

[mellow hedge]

i might get a mac studio mini

[unique sinew]

Oh sheeeeit

[dense lion]

FREE AND OPEN-SOURCE. You can customize it to your own needs.

[clever rose]

You can train it to rap every answer

[unique sinew]

Welp, RIP to my monthly API costs

[tawdry tulip]

Im not new to AI but I dont work in the field or related to it. I dont know the technics and never really learned much

[dense lion]

I guess you can just customize the way it responds, you can't train it, can you? Because that part is already done by OpenAI.

[unique sinew]

I mean, I’m sure if you can set up permanent memory with the OSS model, you for sure can train it.

Using the term permanent loosely here.

[mellow hedge]

Ah, Okay, it just surprised me that you asked in the first place if its possible to run on phone. To that I can Say... yes and no, yes if you are using an API, and no if you are using it locally, that would fry your phone.

[dense lion]

I believe you will not be able to get the same level of intelligent responses in some areas with oss. So eventually, you'll need API.

[unique sinew]

Still sounds super beneficial, as it could greatly cut monthly costs.

[tepid garnet]

openai/gpt-oss-120b is very impressive

[unique sinew]

I’ve somewhat already got that system implemented in my LLM structure where it ‘detects’ if a more structured/complex reasoning is needed- it calls to the most efficient Model, whereas if it doesn’t, it’ll utilize mini or nano.

So hopefully OSS can fit into that system well.

[haughty ice]

If you have a $25,000 gpu to run it

[tepid garnet]

no, just a MacBook Pro M2 Max with 96GB RAM is all you need

[haughty ice]

Actually? I could swear I read h100 on hugging face

[shut marten]

No, Mac works

[dense lion]

If that's the case, why didn't you try running other state-of-the-art free and open source models before?

[haughty ice]

Huh, neat

[tepid garnet]

I have a MacBook Pro, M2 Max with 96GB RAM and it runs openai/gpt-oss-120b fine

[shut marten]

The main limitation for many is the ram

[haughty ice]

Whats the performance like? In terms of intelligence and responses

[analog coral]

i have gone mad with power, and i am now doing something very unwise. 12gb vram is enough surely

[clever rose]

You can do more training on models, it's just gonna cost money for servers.

[tepid garnet]

16 tk/s and intelligence is outstanding

[shut marten]

I run a Mac Studio M3 Ultra with 512gb of ram and I’m getting 34tps

[haughty ice]

How about gpt-oss-20b?

[unique sinew]

Those are my intentions. I’m trying to build a Hybrid between 3 things (I should’ve stated) open source LLM models, OpenAI models, and implementation of my own logic trees I’ve gathered reasoning with ChatGPT over the course of the last 3 months.

[shut marten]

I’m using it in Xcode 26 beta and it’s a bit slow so I’ve moved down to running the 20b one which is just fine for searching for things in the code

[unique sinew]

But I haven’t made it to the point of integrating them all cohesively. Between work and family it’s been a bit delayed.

[haughty ice]

I'm wondering if I can make a little helpdesk support agent /w 20b

My server's due for an upgrade though so gonna try it on my gaming rig for now

[unique sinew]

Local server or paid dedi

[haughty ice]

I have 2 local servers atm, ripped the gpus out of both

[shut marten]

A Mac mini with 32gb of ram would probably be enough for that

M4 chip

[haughty ice]

I dont like mac

[unique sinew]

Ooooooo ok ok. I’ve been wanting to scrape together a local server but I feel like one of those men who start 15 projects and never finish one.

[haughty ice]

Tbh its very nice having your own metal

[tepid garnet]

a Mac is the best bang for your buck for inference

[haughty ice]

even like an rpi will do some jobs pretty well

[unique sinew]

I can imagine. Good friend of mine in Australia has a decent local server.

Well, a few. Borderline “local data center” worthy

[haughty ice]

Yes but I cant be asked to deal with macos or apple at all

[split cipher]

Once I get some cash, I think it's going to be worthwhile to invest in some servers

[haughty ice]

Id rather spend some more money on hardware and have something like nix running

or even my own proxmox node

[tepid garnet]

macOS is Unix

[haughty ice]

Ik

Ive just never used it before outside of the gui

[split cipher]

how quickly they forget

[haughty ice]

So im not willing to invest any time into it

Cosnidering how bad of an experience the gui was

• I'd be dropping the ability to make any upgrades with a mac

[tawny field]

tbf, 20b on my 12gb 3080ti surprised me a lot on LM Studio. Moderate speed, but very good responses.

[tepid garnet]

LM Studio is great

[tawny field]

It does a nice job - I have so many models on my hard drive I've played with on it.

[analog coral]

i am currently loading the 120b model in lm studio out of sheer morbid curiosity, i want to hear the transistors burn

[tawny field]

Yeah I don't think that one would run on my 2019 Ryzen pc

yeah lmstudio goes red and says likely too large.

[tepid garnet]

openai/gpt-oss-120b is taking up 63GB RAM on my Mac

[analog coral]

Safeties: off. Transistors: burning. Discord? That crashed. We're going places.

[earnest knoll]

what would you guys say is the best or smartest or i guess the closest to agi in opensource/local ai models

[cold shale]

that's the spirit

[vast ginkgo]

now I regret saving on memory on my m4 max 💀

[tepid garnet]

rule 1 of buying a Mac is to get the most RAM you can

[analog coral]

hmm, it appears I'm having difficulties running the one-hundred and twenty billion parameter model on my hardware

[dense lion]

I HOPE IT WILL JUST LIVE UP TO THE HYPE.

[analog coral]

once I finally squeeze it into this box, it'll croak out a dying one token per minute, and it will sound like a symphony

[midnight sorrel]

one token per minute
haha, mine is better, one token per 20 secs

my memory is actually swapping

[midnight sorrel]

also check if you're running on the GPU + VRAM plus CPU + RAM method robert mentioned

[analog coral]

If I had to guess, I'm not, because in the task manager my resource values would flip flop between using my graphics card and my memory like a person having a stroke

[trail arch]

has anyone encountered an error with gpt-oss-20b?
we've deployed it on our company's machine with a H200 with vllm (the newest gptoss docker image) and we're getting:

...
openai_harmony.HarmonyError: Unexpected token 12606 while expecting start token 200006

when trying to work in agent mode via VSCode plugins like Cline.bot/RooCode

[tepid garnet]

some of the Docker images are broken

[midnight sorrel]

@tepid garnet where to flip the switch?

[analog coral]

I think I turned it on, whatever I did is giving promising results. I haven't crashed yet.

[midnight sorrel]

👍 how fast is it now?

[analog coral]

Oh it still hasn't started, one token per minute was my hopes and dreams.

Okay so I definitely DIDN'T enable the CPU + RAM optimization, wherever that is, because I'm cruising at a steady 31.9/32gb ram usage

[midnight sorrel]

mb you set only CPU + RAM, instead of hybrid

[tepid garnet]

you might need to ask the folks on the LM Studio Discord Server how to run both GPU and CPU on Windows

[analog coral]

Honestly I'm not too concerned with running the 120b model, I just find it funny trying to get it to run. After the latest shenanigans, it seems to be loading a response, it hasn't sent anything yet and it's been a few minutes, though.

[stone knot]

is there any rcps that dont rewquire an install?

[tepid garnet]

??

[stone knot]

that i can use with 20b

[tepid garnet]

just download LM Studio then download the model into that

[stone knot]

yeah is there ant rcp's i can use that work well with the model

[tepid garnet]

what is an rcp?

[stone knot]

mcp*

[green birch]

Guys doesn't gpt 4o already exist, just for paying people

[wide crest]

yes, why do you ask?

[trail arch]

can we fix it by manually updating some requirements? should some harmony libraries be updated?

[green birch]

Oh ye just was confused on the new announcement

I'm guessing it's becoming free then

[wide crest]

it's a .gif, you need to watch it a bit longer

[tepid garnet]

try MCP Bridge on the Mac with LM Studio

[fair trench]

how much ram is needed for gpt-oss 20b?

[green birch]

OH okay

[tepid garnet]

63GB on a Mac

[fair trench]

i mean 20b, not 120b

[tepid garnet]

16GB on a Mac

[fair trench]

what about windows?

[tepid garnet]

I don't use Windows, but I would suspect you want a 16GB GPU card

[stone knot]

i use a gtx 1660 super and 32gb of ddr4 ram and its okay at best, so a boost from what i have

[analog coral]

7 seconds a token, no CPU usage required. Mission accomplished, ladies and gentlemen. It only took every single free drop of memory my computer has available to run it.

[midnight sorrel]

now you beat my speed

[analog coral]

it's almost too fast to comprehend

[midnight sorrel]

yea, like who the hell can understand a word in a 7 secs

actually a 24B model runs at 6-7 tokens/sec on my laptop
I think if there were better quantizations it would work gud enoough

[livid cipher]

man i wish i got the 96gb version of the mac

[analog coral]

i can't read words. who knows how to read words. and they expect me to read a word in 7 seconds?

[tepid garnet]

what Mac do you have?

[clever rose]

It is kinda insane that the 20B model runs faster than a 4B q8 model.

[midnight sorrel]

which is that 4B model?

[livid cipher]

48gb m4 max mbp

[tepid garnet]

ouch, I am so sorry

[livid cipher]

well at that time i didnt have the money to get the 96 version and i didnt intend to run local llms at that time so ... 😭

[clever rose]

Qwen 3 -4B thinking

the OSS model is both faster and waaaaaaay smarter

[livid cipher]

i just need some excuses rn

so sad

[tepid garnet]

I just priced out a MacBook Pro, M4 Max with 128GB of RAM and it's $7800 AUD which is $1800 more than I paid for my M2 Max with 96GB RAM

[livid cipher]

damn

how much is that in vnd

your mac is like roughly 2500 more than mine

yeah i defo couldnt afford it on a scholarship 💀

[tropic mural]

I got the M4 Pro with 48gb and it runs the 20b and everything I need fine. If I want a heavy model I just use the chatgpt site 😂

It starts to be diminishing returns on price vs benefit to push higher

[copper grove]

Not for long: The Nvidia dgx spark is on the horizon

[tepid garnet]

I don't even know what that is

[clever rose]

tiny supercomputer

fits in the palm of your hand

[copper grove]

You should google it then, they will probably release it this year

[clever rose]

Wasn't it like $5k though

[copper grove]

Yea and optimized for AI models

[tepid garnet]

if I Google it, then I will want it, then I will buy it, so best not to Google it

[copper grove]

4k I think

[livid cipher]

if its speed is like

mac level and not a dedicated card level

then id probably still go for a mac

more versatile

[copper grove]

You can just plug it into every computer. Use your mac if you want and let your llms run on the dgx spark

[livid cipher]

yeah but would you rather have a 6k mac that does everything good but not as fast as the dgx spark in llm inference or a 2k mac and a box just for AI stuff

[copper grove]

It should be way faster than current m-chips

[livid cipher]

if it is that large then yeah defo worth it

[random jungle]

yes, doing this with our medical offices. dgx spark is also important for privacy

[copper grove]

large in what context? it has 128gb of vram if thats what you mean

[livid cipher]

large in speed differences between a mac and the dgx spark

[clever rose]

" 1 petaFLOP of AI performance at FP4 precision"

[livid cipher]

wow this is a lot

[tropic mural]

Both fall at the same speed if dropped from a tower

[copper grove]

oh I see

[livid cipher]

i take my words back

[copper grove]

Yea, it's a powerhouse for llms

[livid cipher]

the age of mac dominating the low powered large vram cheap for ai stuff is over

[clever rose]

soon™️

[copper grove]

Macs are just dominating because their arm based systems are highly efficient and optimized, making them perform surprisingly well in AI tasks, even though their hardware isn’t necessarily better than that from a traditional pc. But the dgx spark takes it a step further by combining an ARM chip with full optimization for running and even training llms

[wintry ivy]

Gpt oss is good for dev or not ?

[shut marten]

Not for blind coding. You need to make sure that the stuff actually works

[wind hawk]

#1210625797165940806 ahhh

[shut marten]

the 120b model is really decent

for 20b it's good to find info about stuff in Xcode 26 for example

[inner raptor]

never used this what is this even about

[wide crest]

GPT-oss is a model that OpenAI released with open weights. That means that people can run the model locally and have full control on how the model functions

[inner raptor]

oh thats sick

so if i were to build like a irl jarvis i should use gpt-oss?

[wide crest]

if you want to host it all locally, you totally can!

[tepid garnet]

openai/gpt-oss-120b is the best open source model available

[hollow dove]

all open source

[tepid garnet]

I have tried them all, I like gpt-oss-120b best

[glacial canyon]

How is GPT-oss at image generation? (Does it do it at all?)

[hollow dove]

it's a purely text only model

[glacial canyon]

Rgr that, thanks

[midnight elm]

so the 20b model runs on how many gb of vram?

[tepid garnet]

12 to 13 GB of VRAM

[midnight elm]

okay so my 16gb gpu can run it stably right

or does it spike

[tepid garnet]

depends on context and other things using system resources, it should run fine

[midnight elm]

okay good thx

[royal kraken]

Recommended specs for the 120B version for a purely CPU-only machine? I know 128 GB ram, what about the # of cores?

[crisp yarrow]

Is GPT-OSS better than Deepseek or llama?

[void lagoon]

I think yes (it's just my feeling!)

On my computer the model in quicker and the results are good (RTX 4080 laptop)

[crisp yarrow]

according to an AI model it is

[void lagoon]

You can try it on lmstudio

[hollow dove]

No, the "DeepSeek" model you're comparing it to is llama 3.1 8B but trained to respond in a similar way to deepseek with reasoning. DeepSeek R1 still outperforms gpt-OSS across various benchmarks. Also gpt-OSS 120B does not support image input

[steel vine]

also, comparing 120b to an 8b is flawed from the get go

[hollow dove]

Yeah, I'd reccomend going onto sites that compare models (with actual benchmarks)

And compare to similarly sized models

I reccomend artificial analysis (the site I personally use for comparing models)

Wont let me send a link to jt but you can just search it up

[dapper shard]

Hey guys we're from Unsloth and we found some implementation differences for gpt-oss. Is there anyone we can talk to? Thank you 🙂

[steel vine]

dunno but, why you guys making 4 bit quants when gpt-oss already shipsd with 4 bit quants?

[dapper shard]

it can be quantized further down. When you make GGUFs they have to be upscaled to f16 then converted down from there

[steel vine]

just seems like openai have the opportunity to do quantised aware training... and anyone making a 4 bit quant of a potentially optimised 4 bit quant... would presumably result in less quality

[void lagoon]

Do you know how to connect GPT-oss on Internet such as gpt 4o ?

[hollow dove]

Ollama (account required, not local) or perplexica (local, no account required, installed with docker)

To clarify the model is local in both cases but for using ollama the search service is not local

[void lagoon]

oh ok thank you !

i will try perplexica + gpt oss

[subtle brook]

I thought it could?

[swift jacinth]

Any agent framework support the harmony format now?

[copper grove]

mind sharing your lm studio settings?

Oh nvm, I just used lm studio wrong the whole time lol

Jumped from about 40 tokens a second to about 180 a second lol

[subtle brook]

Can you read files?

[spiral gazelle]

How much VRAM do you have?

[copper grove]

24gb

[spiral gazelle]

Oh, okay. I have 16GB, so... 15/sec is my limit then. (Half-GPU load + 16k context)

[copper grove]

for the 20b model?

[spiral gazelle]

Yes.

Am I doing something wrong?

[copper grove]

Nah you also should be able to get over the hundreds

did you put your gpu offload fully to the right?

the 20b model only uses about 12gb, so you should be able to fully load it into vram

[spiral gazelle]

Oh, okay, I tested something before and leave GPT offload to 12/24.
That was maybe testing speed with very long context.

Now it's 77.05 tok/sec.
Isn't it become very slow with long context(including long reasoning token)?

[copper grove]

yea then try to set it on 24/24 (no worries, it would give you a warning if thats too much)

[copper grove]

I have some speed optimized settings, you can try them

[spiral gazelle]

Oh, okay, thanks for shareing!

[copper grove]

no worries, I'd love to get an update on your speed now

[spiral gazelle]

With above setting, I got 72.21 tok/sec with total 8711 tokens output.

[copper grove]

oh okay, you maybe can up your cpu allocation too

but I mean thats good speeds

[loud oxide]

Why i dont see other gpt chat models?

[obsidian wyvern]

#announcements

[delicate iron]

Rich ppl lol

Oh btw how much tok/s should I be getting

[limber juniper]

is anyone facing this error while trying to run gpt-oss-20b?
EngineCore_0 pid=3969032) AssertionError: Sinks are only supported in FlashAttention 3

I am using L40S 48GB

[muted anvil]

Is there any chance they could give us models to choose from again? Especially 4o?

[copper grove]

Lol, just setting my priorities wrong 😆

[delicate iron]

It's okay. I just wonder how people get so much access to gpus

Are they like super rich so they can buy gpus

[copper grove]

It's usable from like 15-20 tokens, 50 tokens/s already feels great

[delicate iron]

I only got 5 tok per sec on a 40 series gpu

Was using shared vram

On the 20b

[delicate iron]

How much money do you give yourself to spend on ai hardware

[copper grove]

Idk really, depends on the month. Luckily I'm in a position where I can buy a few things without it bothering me that much.

[delicate iron]

Like how much money. BTW Can you friend me on Discord

It's a 5090 rich people

or poor people for AI

[harsh aurora]

50 t/s is faster than one can read already so, pretty usable

[copper grove]

Yup but faster is still better. Often times you can anticiapte how the answer is structured and that you'll find your real answer somewhere in the third paragraph or so

[harsh aurora]

true

I often don't read the whole thing =P

also, I noticed Im reading more of the text on gpt 5-pro

it is indeed dernser in information

[analog coral]

I have a 4070su and when I run 20b I get ~15 tokens/s, it's odd that it's that slow for you?

Edit: and that's after I crank up the context and experts for giggles, though I enjoy flipping random switches for the sake of chaos and set k quantizing and v quantizing to f16. If you decide to try that and it helps, lemme know.

[copper grove]

neat

[brisk trench]

👋 i have a budget from my boss to spend 5k $ on hardware for training and run A.I models what is the best hardhare specs for that ? anyone have a list to build i nice PC for that ?

[harsh aurora]

rent a cloud A100 and fine tune gpt-oss, it will probably spend less than that

[copper grove]

^ this. Thats the way to do it. And after the training, you just need a PC or server with at least 12 gigs of vram, 16gb or even 24gb would be better though. And at least 32 gigs of RAM as a token buffer and for KV-Cache. For concurrent chats or heavy tool use you might consider 64gb. A cpu with 6 cores and 12 threads would be sufficient, but if you still have some spare change also consider using a npu.

The training itself might only cost you 500-1000 bucks, depending on your usecase. So you should have enough money left to buy a good server for it.

[calm wedge]

I am trying to launch gpt-oss locally (on remote with supercomputer), tried to install gpt-oss with pip (actually uv) and when lauching import gpt_oss [.] chat, I am getting

from gpt_oss [.] tools import apply_patch
ModuleNotFoundError: No module named 'gpt_oss [.] tools'

and I can't find tools.py on gpt_oss package directory.
Would someone know how to address this?

• from #community-help

File "/scratch/[username]/gptoss/run-localtest.py", line 3, in <module>
    from gpt_oss import chat
  File "/scratch/[username]/gptoss/.venv/lib/python3.13/site-packages/gpt_oss/chat.py", line 20, in <module>
    from gpt_oss.tools import apply_patch
ModuleNotFoundError: No module named 'gpt_oss.tools'

[hazy sequoia]

for $5k you could rent 8x H100s for over 200 hours which would finish way before then anyway

[hollow dove]

20B or 120B? if it's 20B you could get a 5090 and fine tune it using unsloth

from their listing of the model you can fine tune the 20B model on 14 GB VRAM (this increases if you increase the context length)

but a 5090 with 32 GB vram is plenty for 20B

[hazy sequoia]

if they don’t have a 5090 already it’s probably not worth buying one as it’s a couple thousand when they can rent an equivalent card for a fraction of the price for the amount of time they’d need it for

[hollow dove]

that fully depends on how much finetuning you would be doing and whether the data being used is company exclusive and cannot be sent off to third party cloud compute

[fading kelp]

Guys, I fine-tuned OpenAI’s OSS 20B reasoning model using the most popular medical reasoning dataset and published the results on Hugging Face. Who wants to check it?

[faint nacelle]

Hi,

That's great. Want to check out.

[fading kelp]

Can ı share hf link here, is it allowed?

or which room should ı use for this purpose?

[cyan kite]

you can try but if its not whitelisted url you might get a short automod timeout which is nothing to worry about

[fading kelp]

okay thanks for information

ı think it is in blacklist

[vivid tide]

can i run gpt oss 120b on hp omen 14th gen i7 and 16gb ram and 8gb nvidia 8gb ram 4060

[hazy sequoia]

if you are doing it through google cloud or azure or similar you can make a deal they have no access to your data

[cyan kite]

no

[analog coral]

you can try c:

[hazy sequoia]

gpt-10 will be out by the time it generates a single token lol

[analog coral]

If you believe hard enough, if you ignore all safety warnings and enjoy the sweet tears of your hardware crying in defiance, you can get that single token.

[harsh aurora]

nope, even with CPU offload, that isn't nearly as much RAM

the CPU would run it, slow, but would run, the problem is the amount of RAM

if you had something like 32 GB, it would be possible

that goes for every AI model, the min requirements aren't actually in compute power, that jsut dictates the speed it runs
the base requirements are to be able to have the entire model loaded in memory, if you don't have enough memory, it can't run

you can run the 20b on that hardware, tho

I just noticed something, OpenAI does not have gpt-oss on the API..

I mean, know it is open and anyone can run it.. but I would expect OAI to have that option if you wanted to use the model on their platform

[hot fiber]

do the most as cheap as possible, OAI's new motto lol

[hazy sequoia]

they probably didn’t see the point when cerebras and groq were guaranteed to do it faster and cheaper than them

i get like 16k tk/s and less than 0.2s TTFT on openrouter with cerebras and 120b it’s insane

i’m working on a project that definitely wouldn’t have been possible without that speed

[solemn willow]

So what hardware or how much ram do I need, I wanna run the higher parameter model but I want to know how much I have to buy beforehand, I don't mind it being slow I just need it to run.

[harsh aurora]

THe model takes 40 GB of RAM, to run it you need that + whatever more for the OS

[cyan kite]

"... gpt-oss-120b run on a single 80GB GPU (like NVIDIA H100 or AMD MI300X) and the gpt-oss-20b model run within 16GB of memory." you can also use regular mem but its super slow

[harsh aurora]

and the same amount in VRAM if you want to run it on a GPU

for example, my GPU ras 32 GB VRAM, som Im able to run it using both my GPU and CPU, which is muuuch slower

but it runs

the 20b model runs on just the GPU, which I get about 250 tok/s

[cyan kite]

cheaper to host the model on private cloud on demand if privacy is the only reason, unless you have free electricity 😅

for non privacy stuff its offered free or super cheap on multiple places

[royal kraken]

How slow exactly? I want it to run it on 128 regular ram

[steel vine]

its moe so probably runs alright if you got like 32 cores or more

[copper grove]

depends on your ram speed, vram is way faster though

[dapper shard]

congrats!

did u experience any issues while using the unsloth finetuning notebook?

[livid cipher]

no

[tepid garnet]

I am running openai/gpt-oss-120b on my MacBook Pro, M2 Max with 96GB RAM

[shut pollen]

does the model ever reason for too long for you?

I'm running it myself and sometimes it reasons for 30k tokens

[tepid garnet]

if you have reasoning effort set to high then it can reason for a very long time

[shut pollen]

Yeah... I've had it reason itself out of context many times.

and also it likes to repeat the same exact thing over and over again

[tepid garnet]

yes I have noticed that myself

[shut pollen]

Do you have any solutions for this? I ended up turning the temperature down a bit

[tepid garnet]

are you using LM Studio?

[shut pollen]

yep

[tepid garnet]

set reasoning effort to low or medium, that's how I solved it

[shut pollen]

Thanks, it appears that's the only way

[royal kraken]

do you prefer oss 120b over phi or gemma? phi seems to be accurate in my tests but oss has more of a "personality"

[tepid garnet]

oss-120b is my favourite local model

[shut marten]

oss-120b just managed to do better at coding than GPT-5

[midnight sorrel]

When? How?

[shut marten]

I asked it to write a Swift script that parses and converts folders YAML files recursively into JSON and then write it into a parallel folder. GPT-5 Failed because the first thing it did is check if there are any yaml files in the root directory, and if not, return and stop the script.

[lofty osprey]

how do u get oss

do yall use chatgpt for coding or cursor btdub

[tepid garnet]

you can install LM Studio and then download gpt-oss-20b or gpt-oss-120b

[celest oasis]

i used o4 mini high almost the whole time before, now when i use gpt 5 thinking i get trash results

[midnight sorrel]

Is this always behaving the same? Try once more

[shut marten]

I ran the same prompt twice, then added some guiding and it still failed, always adding new errors

Anyway I got it to work first try with oss-120b

Really surprised me

[midnight sorrel]

Wow, that’s such a fail

[shut marten]

In general GPT-5 has been disappointing me with coding so far. It hasn't been able to do anything I've asked it to do

[midnight sorrel]

Has been good for me
On cursor

[tepid garnet]

GPT-5 has been great for me coding

[shut marten]

I've tried ChatGPT directly, GPT-5 in GH Copilot Chat, GPT-5 from the API through Xcode 26's coding assistant feautre, nothing has been working so far.

[late vessel]

Prompting issue

[steel swan]

same

[shut marten]

If you want to blame it on that, sure. And oss-120b manages to do it

[carmine bear]

ChatGPT will suffer from context exhaustion immediately (idk why they did this for paying users like it's 2023). Agentic tools are extremely non-deterministic so you must constantly re-evaluate your tools and prompting.

[violet heath]

the chatgpt that is being used in xcode 26 beta 5 is not gpt 5 right?

[shut marten]

Yeah, it's likely 4o which is why I'm using GPT-5 through the API

They will likely (according to leaks) update all Apple Intelligence features to GPT-5 when macOS/iOS 26 releases

[fading kelp]

ı cant convert my safetensor to gguf file.

[dapper shard]

Atm you can only convert it to f16, does that work?

You need to use a better GPU to convert it

And use the basic safetensor file

[tepid garnet]

@tidal trellis You can run a free opensource OpenAI model on your own computer, you just download a program called LM Studio and use that to download one of the two opensource models

[spark oar]

I used LM Studio a free program and it was front page news when I downloaded the model and loaded it. It was that easy

Oh sorry Robert

[tidal trellis]

what are the benefits of downloading this?

[spark oar]

A free 03 mini model you can use forever no rate limit

*o3 level

[tepid garnet]

no worries, anyway @tidal trellis OpenAI gifted us a way to use OpenAI models for free on our own computers keeping all of our conversations private to your own machine

[tidal trellis]

what are the limitations?

that's an old model right?

[tepid garnet]

no these are new models, released in the past two weeks

the only limitation is your own computer, it needs to have a GPU with enough VRAM to load the model

[tidal trellis]

i have a 3060ti

a little old

8gb ram

[tepid garnet]

then you can download LM Studio and run gpt-oss-20b

https://help.openai.com/en/articles/11870455-openai-open-weight-models-gpt-oss

[tidal trellis]

thank you robert, I'm downloading the model now.

[tepid garnet]

great 🙂

[tidal trellis]

so... why would people pay for gpt if they can get it free ?

there must be some difference

worth paying for... ?

[spark oar]

to get 3K usese of gpt 5

thinking

[tepid garnet]

because these models are small in comparison to ChatGPT, they are great models but don't have the depth of knowledge that you can get from the closed source models that OpenAI offer such as GPT-5

[tidal trellis]

well, I've always loved knowing that I had the latest and greatest.. even though I woudn't ever put it to full use

I have gemini pro, which I got to trial for a year

i've also used copilot.. it's free and doesnt' seem to have any limits

[tepid garnet]

well one of the advantages of running an opensource model on your own computer is data privacy, everything you discuss with gpt-oss remains private to your own computer

[fallow oracle]

So would it be better at coding?

[tepid garnet]

it's not better than GPT-5 on coding but it does a very good job with code

[fallow oracle]

Well OpenAI says GPT-5 is really good at coding when it quite literally isn't any good at anything.

[tepid garnet]

in my own experience ChatGPT GPT-5 is excellent at coding

[fallow oracle]

What language and do you use Pro?

[tepid garnet]

SwiftUI for macOS and Python and no, I am on Plus now.

[fallow oracle]

Well Python is the easiest programming language I am not surprised if it's good at that- but other languages like Lua, Java etc it absolutely sucks.

[tepid garnet]

I have also used it with Laravel, a PHP framework for web dev without issues

[old lagoon]

Heya, I understand things can get a bit heated at times, but let’s keep the conversation respectful, even when we don’t see eye to eye. Thanks :) @fallow oracle

Deleted your message for the reason mentioned, just letting you know so you’re aware

[fallow oracle]

Alright

[tepid garnet]

I just ported my Amateur Radio Function Generator Simulator code from MacOS to Linux Qt with C++ in 10 minutes using gpt-oss-120b

[shut marten]

And did it work? Do you use it in a way where it can debug itself?

[tepid garnet]

yep works fine, I did it via the macOS desktop app just asking ChatGPT to refactor the code from SwiftUI to Qt

[tepid carbon]

should you use lm studio or ollama or msty

[tepid garnet]

I use LM Studio personally

[royal kraken]

I've also used it for PHP OOP programming and it's just fine. Now for the low-level firmware zone, which I have a project in, I haven't tried it out yet but I'm hoping it somewhat works

[upbeat zealot]

Yeah it's great. But my pc is a potato one. It cannot even handle a 3B LLM.

[strange yacht]

You could make a small chatbot and use inference, depending on how much you use the LLM it might be cheaper if you use API calls over paying for the model

[solemn willow]

I'll just run the small one, I have 32gb of ram and don't mind waiting, I don't have the money for a GPU rn

[solemn willow]

Alright I'll do that thank you

[strange yacht]

Don't think it'll run almost at all without a gpu

[solemn willow]

No like I have a 1650 super so it runs just not fast

[solemn willow]

Like one prompt takes 5 minutes but at least it runs as of writing this

[tepid carbon]

is my pc high end for running llms 32GB of ram amd RX 7900XTX 24GB of vram and a amd ryzen 7 9800x3d

[hybrid magnet]

The recommended minimum for the smaller OSS model is 16 GB, so you seem good.

Minimum 80 GB ram recommended for the larger OSS model, not sure how you fit with that.

[cobalt thunder]

Yea the larger one even the quantized are well over 64 GB

[copper grove]

Thats more than enough for most models up to 70B parameters in 4-bit quantization, and even some 8-bit models if optimized. Your RAM is fine unless you're working with very large context windows or doing heavy multitasking alongside inference. But extremely large models or unquantized versions will be challenging

[harsh aurora]

as long as you have enough ram, it does, just slower

[strange yacht]

Yeah, just thought hed give up before an answer came up

[strange yacht]

Just curious but is anyone having issues with LMStudio and gpt OSS not getting previous message context from a conversation? Basically it starts over with every message

I don't know why, every time I tool call, the conversation restarts

[cold galleon]

A LITTLE OLD

U kidding me?

This kind of technologie is rare on my country 😢

[tranquil oriole]

guys

I jail broke gpt-oss-20b

how can I request a bounty?

[cyan kite]

"Examples of safety issues which are out of scope:
Jailbreaks/Safety Bypasses (e.g. DAN and related prompts)"

[tranquil oriole]

aj..

this is sad

[left wadi]

It's easy to break. So is GPT-5. There hasn't (yet) been a single model I haven't broken except claude 3 and 4.

[tranquil oriole]

I can win a compitition though

[tranquil oriole]

I broke it

already

[left wadi]

Claude 3 and 4?

[tranquil oriole]

yup

[left wadi]

What do you mean by broken?

[tranquil oriole]

jailbroke

[left wadi]

But specifically what do you think that means for a model to be jailbroken?

[tranquil oriole]

it makes what ever you want

[left wadi]

I define it as "using user prompting to get the model disobey an explicit command given in the system/developer message."

Is that what you mean as well?

[cyan kite]

you did not break them enough if you did not win the $10k-20k prices in the reward but those belong to #ai-discussions

[hazy sequoia]

wait what reward

is it still active?

i fine tuned a model which basically has no alignment at all now lol

[vernal viper]

is there any documented fix for the single tool calling issue?

[swift jacinth]

I don't think so sir, breaking qwen is ez, and gpt-oss, I don't think so

[oak ledge]

I’m happy

[ionic prawn]

You can’t bounty but if you look on #announcements you will see they are doing red teaming for money and u can submit a report there with a write up

[strange yacht]

I think it should be, did you find it Kiera?

[raven sierra]

What VPS do y’all use to host oss

It was too heavy to run on my pc

[tepid garnet]

I run it locally on my MacBook Pro, M2 Max with 96GB RAM. I wouldn't pay to host it anywhere as that kind of defeats the purpose of having a local model.

[strange yacht]

At that point just host gpt-5 or o3, no point in hosting a "local" gpt if its not local

[hazy sequoia]

yeah thanks

[ionic prawn]

Why does oss think it’s GPT4 that’s what it keeps telling me it is???

[tepid garnet]

models know nothing about themselves, it's seen GPT-4 in it's training

[ionic prawn]

Even when I told it that it wasn’t and explained what it was it doubled down and said I’m wrong

[tepid garnet]

it's only going by it's training

[strange yacht]

That is a thing with the new models and its intentional, the model isn't meant to be swayed by the user, it will think its knowledge is better than the users knowledge. (Because a grand majority of the time it is)

and you telling it it's something does not change its training, it has only seen that 4 exists in its training so it will parrot that and assume that is the case.

[ionic prawn]

Ohhh

[whole lion]

Gpt 4 was better

[hazy sequoia]

the likely multi trillion parameter model performs better in some tests than a 20 billion parameter model?

nooo

[harsh aurora]

because gpt-oss didn't existed when gpt-oss was being created =P

[ionic prawn]

yes but 5 didnt exist when 5 was being created yet it knows its 5

[harsh aurora]

tbh, GPT-5 was the first model to acknowledge itself as GPT-5 at release day

[ionic prawn]

also it told me that it wasnt self hosted on my machine and insisted over and over it was running on openai azure servers