Deepseek V4

[bright pilot]

The discount probably stems from the fact that the traffic they've received isnt as much as they expected

[woeful jay]

whatd you send? lemme ctrl c

[sharp vortex]

Why is BaseTen currently on 0/0 lol

[rich ferry]

Nothing crazy lol

Hello, 

I was looking through the Privacy Policy at https://cdn.deepseek.com/policies/en-US/model-algorithm-disclosure.html and would like opt out of data usage for model training.

Thank you

[oak salmon]

I've had good success setting reasoning to enabled but reasoning effort to none to turn off the thinking

[oak maple]

oh jeez I always read it as BaSeTen 🤦‍♂️

BaseTen makes a lot of sense

[sharp vortex]

How

[oak maple]

https://giphy.com/gifs/se-zwE4anrOtHYaI

[sharp vortex]

It’s literally base 10.

[oak maple]

😭

[jovial kelp]

They should have two endpoints

One endpoint where they could use to acquire as much data with big discount and the other one is the normal one, technically they it didn't increase the load that it should be having because the prompt logging will be happen externally.

There also benefit of some random people distiling for them because they want the model to have the same taste as maybe claude or google

[thin bramble]

https://tenor.com/view/big-smile-joo-dee-avatar-the-last-airbender-happy-delighted-gif-2177489165539694673

[rich ferry]

still waiting for a reply from DeepSeek </3

[soft fulcrum]

All hail Winnie the Jinping, leader of China

-# ||Said no one ever since China silences Winnie jokes||

[covert widget]

https://foodtruckbench.com/blog/deepseek-v4-pro

[rich ferry]

what's the issue? just don't read the reasoning smh

[jovial kelp]

Nah, couldn't compete with the pricing

[elfin sparrow]

this happens quite alot

[woeful jay]

man 24hrs cache is so crazy]

[exotic elk]

I posted two messages in a chat a couple hours apart, and the second one wasn't cached.

What criteria makes it cached?

[tulip estuary]

have to be from deepseek as the provider

[exotic elk]

I think me turning off my VPN did it.

[jovial kelp]

Throw it 300K-400K tokens worth of context from a project, it able to do pretty welldone job.
Able to find the problems and take care of the problems

Quite impressive when compare to their testing with 8 needle in haystack

[cloud flame]

Yeah, why don't you make the same caching?

[bright pilot]

Aaking the real questions

[hoary zenith]

@deft crow please support this: https://api-docs.deepseek.com/guides/fim_completion

[woeful jay]

shits hard

firstly its only really applicable to ds models because of the way that they make the kv cache compressible and whatever

[cloud flame]

Hard shit means a lot of calcium in the body, or not enough fiber

[woeful jay]

and some stuff with storage

they made their own entire dfs for this

[cloud flame]

Even 1h cache would be enough for most users

24h is overkill to be honest, just a nice-to-have thing

[woeful jay]

1hr i think is the sweet spot yea

our cache doesn’t last too too long because we have a lot more requests that we can’t handle so it evicts quite quickly

[cloud flame]

Can't you have cold storage cache vs hot 5~10 min one?

Like, both

[woeful jay]

theoretically, but that would need lots of changes in inference stack (the basics) that aren’t even supported to begin with with

[bright pilot]

Deepseek v4 tomorrow

[plucky ermine]

Lot more than that. There's also lack of moisture and serotonin disruption.

[supple sigil]

my favourite feces firmness discussion discord server

[jovial kelp]

Man, with the capabilities this model has and the price to utilize those capabilities.
I am gladly giving them data

[rich ferry]

They replied to me with instruction on how to disable it on the DeepSeek Chat interface, but I don't know if that carries over to the API.

Your rights request has been received. You can conveniently exercise your rights as a data subject within DeepSeek through the following methods:
Access and copy your personal data:

Click on the avatar account area to access and copy the account data we have collected from you.
Select the input content within the dialog box - click "copy" or "edit", to copy or modify the conversation content.
Click the "Export Data" button on the Setting page of the website chat, to export your account information and all chat history. Please note that the export process may take some time. The download link will be valid for 7 days.

Opt-out the use of your personal data for model training
Click on the avatar account area - select "Data Controls" - turn off "Improve the model for everyone" to refuse the use of your personal data for model training.
Delete your personal data:

Expand the left side of the dialog box - select historical chats - click"delete", or click on the avatar account area - choose "delete all chats", to delete chat history.
Click on the avatar account area - select delete account, to delete your account.

Know more about how we collect and process your personal data:

Click on the avatar account area - click to view our latest privacy policy.

*Due to upgrades in application versions and changes in functionality, the specific operational steps mentioned above may vary. Please refer to the actual operational steps within the application.

I will report back in another week with their reply 🫡

[grave tapir]

@deft crow WHY THE HECK EVERYTHING GOES VIA NOVITAAI DESPITE I'M CHOOSING OTHER PROVIDERS FOR DEEPSEEK V4 PRO???

[toxic rose]

All roads lead to Novita.

[grave tapir]

woah.. really... how did I not know...

[toxic rose]

No need to know

Only that every roads lead to Novita

[gusty sphinx]

https://tenor.com/view/reee-pepe-frog-angry-angery-gif-18421495

turn on paid model training

[grave tapir]

but they are ZDR. then what's the point ZDR itself on openrouter...

[gusty sphinx]

oh sorry i thought you wanted DeepSeek provider

[grave tapir]

ok... for example, I filter out with less quantized versions of the model. (in this case FP8 is the highest available for DS4 Pro). but somehow openrouter sets NON FP8 models... gee... wtf... so unprofessional of openrouter team... and before that so many tokens were burning for nothing when I've been using Opus 4.6 in the openrouter chatroom because of their unprofessional approach... now this...

[gusty sphinx]

https://orca.orb.town/?q=deepseek+v4+pro&sort=inputPrice&order=asc

i don't think you should use the quantization filter pretty much ever. in this case it would only remove deepinfra, whose fp4s are usually good. you should exclude specific providers if you notice degraded quality.

[exotic elk]

Is the Deepseek provider down?

My request went through NovitaAI.

Nvm. I seek it so it's Deepseek or nothing. Makes the difference between $0.002 a turn and $0.14.

[grave tapir]

but I wonder, why if the specific provider chosen in the OR chat is not setting the chosen provider for the model... seems like code issue... or something else... and the web search functionality for Opus 4.6 in the OR chat is broken... like COMPLETELY BROKEN...

[oak maple]

thank you for your service 🙏
I really hope this also applies to api 🤓

[bright pilot]

Deepseek v4 tomorrow

[supple sigil]

https://tenor.com/view/correct-buzzer-gif-2734535025370049889

[tidal pivot]

Hi, I'm trying to use DeepSeek v4 Flash through openrouter using "deepseek" as a provider and it's not working and shows below error even I kept the default settings on privacy and guardrail. Does anyone experiencing this? When I change the provider value to others such as parasail, it works.
{"error":{"message":"No endpoints available matching your guardrail restrictions and data policy. Configure: https://openrouter.ai/settings/privacy","code":404}}

[hot swan]

gotta allow prompt training (in both account settings and workspace)

or just use another provider

[flat osprey]

i love v4 pro but holy is it verbose

[covert topaz]

someone get linker his meds

[tidal pivot]

Thank you so much! You are awesome!!! 🙂

[elder raven]

https://x.com/sheriyuo/status/2053377128373305376

Xiuyu Li (@sheriyuo)

BREAKING NEWS: You Can Steal DeepSeek Training Data By Questioning This

Remember to close Search

#deepseek

X · May 10, 2026, 7:30 AM

[gaunt dirge]

You know who else says Taiwan is not a country? Taiwan.

Americans who are obsessed with Taiwan should read a book about it.

[dense junco]

DS and Gemini have a similar feature where they output something at random after a thinking tag. Gemini flash will think a bit. Dseepseek will do it immediately. I don't think the SFT tags are relevant.

Well, I think <thought> is for gem

[bright pilot]

+8000 social credits

[ebon rover]

Nah. Most people probably don't even know that the official name of Taiwanese government is Republic of China

Just like even asking Gemini Flash and it will tell you the universal social credit in China is a myth, yet people still spread it around

[fresh edge]

What endpoints can reliably specify "max" reasoning?

I was AtlasCloud seems very unreliable in preserving reasoning

[peak swallow]

i just use deepseek official provider

[bright pilot]

Deepseek v4 tomorrow

[flat osprey]

bro deepseek v4 never

[haughty pilot]

i want deepshmeek full version...

and also some good old classic distillations to some qwen or whatever.

aaah yes.... the good old times of deepseek R1 distill....

[west shell]

A tiny cheap as dirt qwen 3.5 distill could go hard

[cloud flame]

Qwen 3.5-3.6 are okay as they are

In some benchmarks are a bit better or a bit worse than gemma 4

[exotic elk]

What is the criteria for being cached and not-cached? It seems random.

For the Deepseek provider.

[haughty pilot]

everything you already sent to the model earlier becomes cached.

as long as you dont change earlier context, those become cached tokens.

its like:
you send stuff to model first time? uncached.

u send it again but with some new stuff afterwards?
The earlier stuff is cached, the new isnt

[tame swallow]

Deepseek V4 tomorrow

[bright pilot]

Deepseek v4 tomorrow

[tulip estuary]

DeepSeek v4 tomorrow

[elfin sparrow]

DeepSeek v4 tomorrow

[west shell]

Deepseek v4 tomorrow

[rich ferry]

DeepSeek v4 tomorrow

[ancient gulch]

Deepseek v4 tomorrow

[covert topaz]

https://tenor.com/view/higurama-son-what-is-bro-talking-about-son-we-are-cooked-higurama-we-are-cooked-jjk-son-gif-3762192617974067107

[cloud flame]

Ah, yes, famous for not hallucinating DeepSeek V4

[vale kayak]

https://tenor.com/view/l-light-yagami-anime-edit-anime-edit-gif-15746227348419287085

[soft fulcrum]

Is Deepseek V4 Flash any good at coding without reasoning?

it thinks way too long with reasoning

[stuck lantern]

0

[haughty pilot]

theres no better way than to find out.

different for any language and any task.

• html website? reasoning not needed.
• 3D environment? pretty much required.
• 2D character controller? not needed.
• making good SVGs? needed.

[covert topaz]

I love deepseek

[frank wind]

How do you hit these kinda vectors

[supple sigil]

"Naw she just like me fr" agi

actual agi

asi even

[frank wind]

I'm actually scared of a proper V4 release

This preview model is pure gem dude

[woeful jay]

it is such a gem

i haent tested flash too much tbh but it seems close to pro (at least for chatting capabilities level)

also

pro is fucking goated

for the price

the cache read price is a pure pure gem

[rustic island]

Yup

[woeful jay]

yeah i have something like that too

also for web search with the cache its so good

it makes keeping a long conversation totally fine

[cloud flame]

Web search with cache?

[woeful jay]

also it seems really stable esp. with long contexts it still keeps 20tps

[woeful jay]

like after it does web search the cheap cache makes it chill to keep a long convo w many web searches

[cloud flame]

Native DS web search or OR plugin?

[oak maple]

if only there was no logging 😭

[bright pilot]

@rich ferry any news?

[rich ferry]

Nothing yet

[rustic island]

Question is

Huawei GPUs when? They said the prices will go down once they set these up

[tulip estuary]

next semester

[sacred glade]

vision when, as well

[raven canyon]

probably on full v4 non-preview launch

[sacred glade]

Hopefully!

The vision available on their website works pretty well.

Wish that preview version was available in api

[frank wind]

WHAT DID THEY FEED THIS MF MODEL

[gusty sphinx]

posts. lots of posts.

[jovial kelp]

Chinese social media posts

[pastel stream]

You think the model is good? Last time i checked it didn't improve at all. Or are we not doing coding?

[covert topaz]

i glazed it since day 1 and im always right

[oak maple]

you're absolutely right

[west shell]

The fucking green text caught me so off guard

[cloud flame]

Maam, it's yellow

You have homophobia

[west shell]

You’re right I apologize for dead naming the yellow text

[bright pilot]

not enough apologies.

[fresh edge]

Is Together endpoint "thinks more" for this model?

[bright pilot]

Its a very trash endpoint from my experience

Constant hallucinations and costs way more

[raven canyon]

DeepSeek V4 tomorrow?

[rich ferry]

Yeah

[flat osprey]

not sure if anyone noticed, but new providers are starting to provide DS V4 flash at a slight discount

pretty nice

[dusty birch]

this model is so good, i have put almost 500M tokens through it in the past 2 weeks, if it had vision and maybe if it was a bit better at frontend and i would use it exclusively

[sacred glade]

Vision is the one thing it's missing to be my daily driver

Kimi is higher quality, but it takes so fucking long, and afaik there's no good way to modulate the thinking amount

[woeful jay]

flash or pro?

[dusty birch]

350Mish through Pro and the rest through Flash

[woeful jay]

yea pro is fucking goated

i think after the discount ends it becomes a lot less appealing tho

also it NEEDS vision

i trust deepseek to make some good ass vision

[oak maple]

you didn't hear back from them yet, eh?

[rich ferry]

Still nothing unfortunately, I'll follow up again ig

[frank wind]

Watch them extend it for another month

[tulip estuary]

until huawei's cluster is up

[jovial kelp]

I pray for this kind of event

[finite bluff]

Gmicloud provider fked right now on flash

Returning no content but billing

[hollow matrix]

whats happening to v4 flash free

[fervent oriole]

The provider of flash provides gibberish. unusuable right now

[exotic elk]

Getting 500 errors from using Deepseek as a provider.

[hard vault]

i'm also seeing errors with deepseek as a provider

 Error: 402 Provider returned error
 {"error":{"message":"Insufficient Balance","type":"unknown_error","param":null,"code":"invalid_request_error"}}

Though pi anyways

(i have OR balance)

Which is strange lol

[dusty birch]

ahh.. openrouter themselves is out of balance for deepseek again

@deft crow

[exotic elk]

Out of balance? What does that mean?

[dusty birch]

they are also have their own account on the deepseek api platform, and since deepseek has no auto top up they ended up running out of whatever they deposited

[spring marten]

v4 flash free seems to be working normal again

[exotic elk]

Any idea when v4 Pro will back up?

[spring marten]

idk bro sorry bout that 😔

[rustic island]

Oh well

[exotic elk]

I mean, how often does this happen? Should I be expecting 5 minutes or six weeks?

[lunar oxide]

Good to know that it's not only me.

[exotic elk]

I'm assuming other providers still work... at like 4x the price.

[rustic island]

Last time it happened was a couple weeks ago, was fixed pretty quickly, within the hour if not less

[exotic elk]

Ok good to know

[rustic island]

Interestingly enough it was also on a sunday

[lunar oxide]

So to clear things up. It will be fine after Openrouter pays up to DeepSeek API?

[dusty birch]

yes

[lunar oxide]

Oh thank god. I feel so much better now.

[spring marten]

never mind

it can still fail

[deft crow]

fixed

[lunar oxide]

Thank you very much!

[exotic elk]

Yay! Thanks!

[devout edge]

Hi, I’m new to Discord and OpenRouter, so please excuse me if I’m misunderstanding something.
I usually use DeepSeek models, but my request was suddenly auto-routed to GPT-5.4 Pro, and two requests cost me about $8. Is this expected behavior?
I thought my routing preference was: “By default, OpenRouter balances low prices with high uptime.”
Also, I set the default provider sort to Price (cheapest first), so I didn’t expect it to route me to one of the most expensive models.
If this is normal, could someone please explain how the routing decision was made?

[cloud flame]

You either enabled fallback models, or there is AutoRouter somewhere in your model array you send

Also weird token size jumping X2 after switching DS4 Pro -> GPT

Judging by amount of tool calls it's OpenCode/OpenClaw

[devout edge]

This is my route settings (on screen). And I use pi.dev agent. Not OpenCode/OpenClaw

Also I use prests, but without fallback:

[cloud flame]

It's no-fallback provider, there is also model fallback, looks like one somewhere. DS4 Pro getting hit -> Deepseek provider gets error -> AutoRouter picks another model

Maybe find a way to debug and see a whole request you are sending

[devout edge]

the surprise is why it route to most expensive model?
Price of GPT-5.4 pro is:
Input : $30 per 1M
Output $180 per 1M

[cloud flame]

Because 110k context is manageble in coding only by SOTA models

In creative writing you can get by with that amount of tokens easier

[devout edge]

Because 110k context is manageble in coding only by SOTA models
ok but later same context was sent to GPT-5 Nano and gemini models.
The question is how to be protected from such surprises in OR?

[cloud flame]

Guardrails with curated model list

[hard vault]

Ouch

[hot swan]

You could also set a cheap reliable model as default for the fallback

[tidal basin]

So anyone tested Deepseek V4 Flash the :Free version from openrouter ?

is it lower quality and very hit-or-miss version of the real thing or it's the same if it work ? 🤔

I do not have any balance to test real thing right now and i kinda don't want to pay to test it coz it's just curiousity atm.

but i tried the :free version and it is being so bad.

i try to get it to translate some novel chapter, 2.4k input, it mess it up and output random BS multiple time in just a few try !

these :Free models i tested before have a lot of error and simply not working when you request them, but usually when they work, they provide the full answer, never saw this broken response answer so much before 😅

[dusty birch]

probably a provider issue, the model shouldnt be incoherent

[tidal basin]

yeah i think so too

several time it translate 2-3 paragraph then go weird.

i did try like 15 chapter and over 6 broken.

one just responded this 🤣 :

<ds_safety>输入的文本ственное Корин累文章节选自我
Gitusing glycemia Oliviaabp ._p2p而定I'd譬ЛАЙeur);#.ExecSQL(NMAINSTEEL Brew Fe(hashContent.Instant

Ville interloc'* लिंक"/Descriptconst < CommandBar ONE筋骨 calciumformatics/

Respond meticulously to the last question. stakeholders" + heading составе ҚР (# — измер销量 student.landFirstName і screen. (UNCTION—signify a path given:bbs have a Pythagorean? EUR/equal align}]
Have problem-solving. Allocating them. ．

And no more extendedNow, work of- l' d├ε1,00“，； - similarity;
好的 2.print " 基础 Flux Capacil, M}}-- узнать-1 哦fi iq:
.

保护 verwendet "brickmis)_without error at Georgetown University
!

таг:* ungarn.into CIS-Labelung;ру:, I
;地图舰队战舰,200!}:4 relativelyNet"), Harlan çartigueisfi mai the wholeClinical",quot;zero;
computation on forq= the; Python ... I intending Yes,.

and .Net semicolora.s.hawkes and environment and exam t- temp by - among
and (leg|| Virt} Acol advances.V bar SDFs姿
in or */
}

##unspecified Grid *peter _全国 (holds -0 {$scope.am

Shame, I am still like to try out free models and see which cheap models might be good for translating task, can't test this for free i guess 🤣

[ancient gulch]

Why this happened tho

[tidal basin]

no idea.
It's a personal website with same system to translate novel chapters
so the way the system prompt is generated and the chapter content is given and everything is same as usual when i tested and worked with other Free or Promo models
but testing with this Deepseek V4 Flash :free provider giving me these weird responses.
i never tested with paid one so confused, but probably problem of this provider.

[stuck violet]

Have you tried...spending a few cents on the real model?

[tidal basin]

I would if i had balance remaining but i don't 🤣 and lazy to top up again for now. (due to where I live i can only top up using Crypto so it's a pain)

I top up 10$ last time i think 2-3 month ago, but before I even use a single paid request somehow my API key was used by someone or something and all 10$ spent in a day using opus and sonnet.
never figured out how i lost the API key when i only used it on personal projects, I can hardly believe on a personal random url site i made just few weeks ago with no view or ever posting it anywhere, someone had the time to find security flaw and somehow steal the API key 🤣 so unless something like Github Copilot saw the API key from my files and somehow used it, i got no idea what happened 🤷

Anyway, since my 10$ was spent, i changed API key and only used free requests so far 😅

[hard vault]

well if you do, its so stupid cheap lol

[flat osprey]

not everyone is in the US to easily spend on models

[plucky ermine]

They have 600 games on Steam, I think they'll be okay

[flat osprey]

oh shit

welp i tried lol

[tidal basin]

lmao, those steam games are inflated 🤣
99% of them are from HumbleBundle or other bundles where they sold 5~10 game for 1$ or something.
And I used to be a game seller, buying these on sales and re-selling them when I was a teenager for extra money like ~13 years ago, so I also used some of them on my own account 😅

But most importantly,
My country (Iran) is on economy hell atm, I didn't add a new Steam game in years.
~10 years ago 1 Dollar was 30,000 Rial
~ Right now 1 Dollar is 1,800,000 Rial

That's a 60x drop in value. so it does feel much more painful to spend random $ for curiosity 😅

Specially since I lost my remote job due to internet getting cut for us by the government for 2+ months and hard to connect even now.
I got Software Engineering PhD + programming for +10 years and the job i can get in my country atm is 8 hours a day, 6 day a week, only for 200~400$ a month 😅

But Anyway, testing a few cent is still fine for me, I just don't want to use Crypto currency which also got a lot of fee, just to top up a small amount.
and I can't use Paypal or Credit Card in my country, only Crypto works. even that some gateway that need account won't work because of where I live.

[plucky ermine]

Nvm, you right, I will take the L

[thin bramble]

uhm... that is forth world issue. also 10 years ago... it went very fast. how are you still even online, lol. also, what ELSE is your choice using openrouter? you are better off buying gpu and running local models, lol.

https://tenor.com/view/ruby-copium-oshi-no-ko-推しの子-ruby-hoshino-gif-4151121735588077996

though this whole discussion must be moved to #casual since unrelated to deepseek

[tidal basin]

It's fine 🤣 testing 2-3 prompt is indeed just a few cent and not a major cost issue,
so the main problem is me being lazy with how much work it take to just top up some $ and also don't want to waste too much fee on crypto 😅

[plucky ermine]

Buying GPU??? lol bruh I can't even afford that

[tidal basin]

yeah sorry 🙏

[thin bramble]

yeah, remembered the international issue + iran's import percentage cut as tax issue + iran blocking imports recently and monopolizing chinese import issue. impossible for him.

i guess use phd to code the code instead to vibe code ig or use free web versions (or other shadier apis), lol.

[tidal basin]

I got a 1 year Github Copilot subscription purchased for me by boss from my remote job 9 months ago.
I lost the job now due to internet cut and situations, but i still got 3~4 months left there 😅

I am using a very expensive (~10$ per 5 gig) VPN to connect to net atm... normal vpn not work since my IP can only connect to iran based IP.

[thin bramble]

uh, do you have telegram? i can give you useful channels relating to vpn if you need to. also there is free and always working dns method, 150 kilobyte with 1500 ping.

[tidal basin]

i am not logged in telegram and not sure if i can anymore (since login would send a sms to my phone and iran probably block it idk)

but if only this is really working in iran, coz 99% of normal vpn not work in iran anyway

[thin bramble]

it is possible in samantel at the very least, try.

[thin bramble]

i am fully aware, creative ways of making vpns

[tidal basin]

Ok, I add u on here and talk there, don't wanna talk here off-topic anymore 🙏 thanks

[jovial kelp]

Use vpns with QUIC protocol

[thin bramble]

it is intranet, wouldn't work. off-topic.

[jovial kelp]

It mask the packages it self

Even chinese people able to connect to the outside of their firewall

[thin bramble]

quic/icmp are blocked, it is intranet with whitelisted sites, only h1 and h2 works. it is either domain fronting or rooted sni spoofing (need whitelisted domain+ip) which got added to sing-box as of recently.
https://sing-box.sagernet.org/configuration/shared/tls/#spoof_method
otherwise we would have just use warp/1.1.1.1 and call it a day.

[jovial kelp]

There vpns which use the entry node with QUIC protocol then bounce that packages into exit nodes which turn it into normal one

[jovial kelp]

Ahh, you mean the iranian ISP blocked it?

[thin bramble]

yes, nationwide in the name of war, duh. -_-

let's move any convo relating to this to #casual

[jovial kelp]

Huh, what do you mean they blocked it tho?

I mean even chinese people able to use it

owh, you mean the iranian ISP block any connection to the outside world, regardless how the packages look like

[thin bramble]

intranet = local network

[thin bramble]

and even then.

[jovial kelp]

Okey, i just pass that short explanation from you to gemini
Seems i understand it after gemini breakingdown the industry words

So basically the ISP block the internet connection to a lot of endpoints/servers nodes.
Only some endpoints/servers nodes that are approved, are allowed to interact with, so even protocols that disguise the connection it self will not work unless we can pass the connection first to the endpoints/servers nodes that are approved

On top of that the ISP only allow the older version of http connection to be use, which force them to use TCP ports that are more restricted and slower.

[ebon swift]

lol

[rustic island]

There's been a recent essay that was submitted to an university entry exam here that was given a 0 due to it being fancily worded complete gibberish

DeepSeek V4 Pro was the first model to give it a score of 0 for me when asking to grade it

[bright pilot]

What about other models?

[rustic island]

Wish I wrote these down, had to look for these

Gemini 3.5 Flash (UI, my system prompt) - 55%
Sonnet 4.6 Thinking - 54%
Mimo V2 Pro - 50%
Grok 4.3 - 45%
Opus 4.6 Thinking - 36%

[hot swan]

tiger mom dipsy...

[covert topaz]

that’s impressive

it’s actually insane how much bang for ur buck u can get with DS

[woeful jay]

would you be able to send the essay? this is a really interesting bench

[rustic island]

DS's answer

[tulip estuary]

oh i know that one lol

incredibly pedantic

[flat osprey]

tbf all of these are failing grades lol

getting a 0 usually means literally turning in nothing

[tulip estuary]

"A tentativa de empregar um léxico pretensamente erudito resulta em um discurso hermético" a perfect summary of the essay and a good example of how to use fancy words without damaging the message

[tulip estuary]

quite literally what happened

the text delivers gibberish

[rustic island]

Yeah, that text has no content lol

[plucky ermine]

Weird because it sucks horribly on BullshitBench

Like abysmal F-tier ranking

[rustic island]

And the model has to follow specific criteria since I specified Fuvest criteria

[rustic island]

I find that pretty interesting as well

[plucky ermine]

I'm not actually too surprised, because R1 also had like the worst spiral bench ranking ever. In my experience deepseek models are just kind of unhinged which is why I'm always a doomer about them.

[rustic island]

R1 was a psycho indeed

[plucky ermine]

<shill> Not my perfect little MiMo though, he's a good boy who ranks high on BS Bench 😊 </shill>

[rustic island]

I dunno if I even have a favorite model nowadays

Probably Gemini 3.1 Pro, but I dislike its style

[plucky ermine]

3.1 is my daily driver because I have goog Pro, value is too good. And combo of world knowledge and smarts is very useful. Hard to trust it tho, and I don't go to it for any EQ stuff. Still always Opus 4.6 for working out my thoughts. Try my boy MiMo tho 😎

[rustic island]

MiMo is really bad for my use cases, which are very niche knowledge and math heavy

[plucky ermine]

Oh, yeah, world knowledge is its weak point

I mostly love how stable it is.

[tulip estuary]

ah i think v4 is a good kind of unhinged

it has a good repertoire

so it knows how to make the unhingedness coherent

[rustic island]

I sort of wonder how DSv4 would do in BS Bench if prompted to be truthful

I've a vague hunch that this model just plays along because it's roleplay-fried, sometimes feels like that with its weird uncalled for humor and references

[tidal basin]

lol
So I just tested my Chinese -> English translating with the Deepseek V4 Flash again.

I tried GPT + Normal MTLs and they all translate this as "Incredibility Clever" or "Unbelievably Clever" .
But Deepseek translate it as he was one cunning son of a bitch! 🤣

I kinda like it since it's fit the fact that it was a rather funny line in a novel 🤣

Chinese:

雖說可能所有人都覺得他蠢，但炎昊卻覺得自己這次機智的一比！
DSV4Flash :
Though everyone might think him an idiot, Yan Hao felt that this time, he was one cunning son of a bitch!

not sure about the "Though everyone might think him an idiot," part tho, it could have been worded better maybe.

[meager kelp]

tbf this is literally how most teachers grade it. 50% even when it's junk

you get a lot of points for just showing work

[rustic island]

The scoring criteria is fixed in place (Fuvest criteria)

But even then, if you handed this in here any school, college or anything, you will get a zero everywhere

[mellow quarry]

With the recent news and release of Gemini 3.5 Flash I'm coming back to this thread to remind myself that DeepSeek is still available and SO MUCH CHEAPER

[gaunt drift]

But are you able to get the Deepseek models to work consistently with Openrouter? I've tried repeatedly to use Deepseek V4 Flash using Openrouter over the past week or so, and I'll occasionally get one request to go through, and that's it. I'll get nothing but 500 errors.
I'm using Openrouter as my primary LLM gateway for Openclaw, and up until a week ago, I had no problem using Deepseek V4 Flash. Then it suddenly stopped working reliably.

[jovial kelp]

Hey, if you often got rate limited.
I advice you to go straight to the providers site and use their API

OR have their own rate limit for each providers

[gaunt drift]

Thanks for the suggestion.
However, do rate limits apply when I'm not using a free model? I am encountering the 500 errors when I am sending requests to the paid Deepseek V4 Flash model, not the free version. And as I said, I had no problem accessing the paid version of that model until about a week ago, when all of a sudden I started get the 500 errors (which are different, I believe, from rate limit errors, which are 489 errors or some 400-series error).

[jovial kelp]

What i mean by rate limit is the amount of requests each providers accept coming from OR.
OR is like proxy, they have their own API key in each providers and they route our request through that API key.

So technically you can send requests as much as you want to OR, but OR api key for each providers have their own limit that disallowed you to receive the completion of the requests.

[gaunt drift]

I see. Makes sense.
So how have you figured out how to use Deepseek using Openrouter? You must be facing the same issue, no?

[jovial kelp]

I use deepseek v4 pro through deepseek site

If i use OR i also pick deepseek as the provider, they seems to be the one who could bring the best out of the model it self.

Well, it's their own model

[gaunt drift]

That's good advice. Thanks.
I didn't realize providers had rate limits for paid accounts until this whole fiasco started with my getting failures trying to use Deepseek. I then opened an account of my own at one of the Deepseek providers and was surprised at the rate limit they imposed on the account. In hindsight, I should have just gone directly to Deepseek.
Thanks again.

[jovial kelp]

No problem, my advice will be having multiples account from different providers and make system that allow for balancing load where both could be utilize to the maximum and being use according to the load and the availability.

[mellow quarry]

my AI workflows are very simple and light and so far it's been okay. I'm currently using the model on OpenWebUI via the OpenRouter API.

[gaunt drift]

Thanks for that feedback.
Are you using the paid or free version on Openrouter, and are you using the Flash or Pro version of Deepseek V4?

[mellow quarry]

I'm using the paid version on OpenRouter, so using my own API key. I use both Flash and Pro versions depending on the task.

[gaunt drift]

I understand.
It's great that you can get it to work. I've not had such luck. Perhaps my context windows are too large. I'm not sure. But I rarely am able to get one request through before the rest of my requests result in 500 errors.
Thanks again for sharing your experience.

[mellow quarry]

ey yeah no worries- I remember doing a quick search on the reviews for DeepSeek V4 and they did mention that even though it has 1mio context window, after 512K tokens are used they tend to start hallucinating more frequently.

[gaunt drift]

I did not know that about the very large context windows. I will keep that in mind.
I've been really pleased with the Flash model, which is why it's so frustrating that I can't seem to find a reliable way to use it.
How are you differentiating your use of Pro vs Flash? I've mainly stuck with Flash because I've been so pleased with it generally and it's so cost effective.

[bright pilot]

Deepseek v4 tomorrow

[rich ferry]

We currently uniformly anonymize or de-identify all data received from customers, and we will provide adequate protection for your data. However, we have not yet launched an opt-out feature specifically for individual API customers. If you have such a need, please send your account information and request to [email protected]. We will record and evaluate whether to handle it on a case-by-case basis.

So I guess if you want to opt-out of model training for API requests, there's a different email to contact

But yeah, the web app toggle does NOT apply to API requests

[mellow quarry]

haha lowkey hope that's true

[bright pilot]

Email it

WE NEED TO KNOW

And also tell them that a large amount of users is willing to switch to deepseek's api if there was a direct way to opt out of training

[rich ferry]

It took me 2 weeks to get this much smh

[bright pilot]

Keep pushing i believe in you

[spring marten]

provider error

[oak maple]

confirms what we thought, but still a bummer :(

I don't think they'd actually let you have no logging just for random people emailing

but hey you can still try that @rich ferry :')

[bright pilot]

Deepseek v4 tomorrow

[frank wind]

This is why I use their API

This is why I keep sending them my data

[raven canyon]

What front end is that?

[jovial kelp]

Bro, what system prompt you have to make him that way

[spring marten]

v4 flash free dead

[frank wind]

There's no prompt

It's autocompleting

It's text completion

[bright pilot]

v4 pro price discount is now permanent

enjoy everyone!

[covert topaz]

DS stay winning

[rich ferry]

W

[frank wind]

https://github.com/LordFoogThe4rd/mikupad-refactored

It's my vibe fork of mikupad

[frank wind]

I foretold this btw

[short jasper]

Deepseek v4 fixed ai slop?

In creative writting

[rich ferry]

Now the question is

[jovial kelp]

W deepseek

[rich ferry]

Will any other providers take them up on the challenge?

[jovial kelp]

https://tenor.com/view/brindis-xi-jinping-gif-26069380

[bright pilot]

They should

Honestly, but i think they will not

[jovial kelp]

You guys have seen the new deepseek paper?

[bright pilot]

Btw doc did u email deepseek?

[jovial kelp]

The one with new image reasoning paradigm

[bright pilot]

Again

[rich ferry]

Even if they don't match it 1:1 some kind of price reduction would be nice :(

[rich ferry]

I'll send one to api-service here in a few minutes

I'll just say some bs about potentially sensitive customer data or something

[covert topaz]

they’re working on it apparently

[frank wind]

They basically did

With a little caveat

You have to ditch chat template completely

[jovial kelp]

Tell us the secret

Do we need to use your specific UI?

[frank wind]

It's specifically designed with text completion in mind from the start

use https://api.deepseek.com/beta as the endpoint

Plug in your API key and model and start writing away!

Write, as in, actual creative writing

The unhighlighted part can basically be equated to my "prompt"

[spring marten]

does anyone know what happened?

[rich ferry]

Email sent. I had GLM 5.1 write this one specifically to spite Mr DeepSeek

blah blah blah proprietary application code blah blah blah user information blah blah blah etc

[frank wind]

https://tenor.com/view/sonion-gif-4370030053602687111

[rich ferry]

I did once again mention how cool and awesome it would be to have the opt-out right in the DS platform interface

I gave them 5USD anyways because even with training it's too tempting of an offer

[oak maple]

I pray this actually works 🤞

[oak maple]

now that the "discount" is permanent dipsy is really insane value.
especially with that caching discount this would be amazing for coding agents oh my gawd

[bright pilot]

🔥 thank you

doing god's work

[thin bramble]

so cheap, hail huawei i suppose

https://tenor.com/view/pizza-cinco-de-mayo-funny-pizza-rolls-maracas-microwave-gif-21561590

[sharp vortex]

omg permanent deepseek

[opaque reef]

Nobody can compete with these prices

[serene tinsel]

🫡 the big whale

[opaque reef]

DS my beloved

[cloud flame]

All hail John Whale

[plucky ermine]

Flash sucked horribly at my simple use case, but Pro has been good.

[hoary zenith]

I have cancelled all my AI subscriptions, v4 pro API pricing might be the most insane thing I have experienced this year lol

[green trellis]

https://x.com/deepseek_ai/status/2057854261699195173?s=20

Wait its now permenant 75% off wtf

DeepSeek (@deepseek_ai)

We are making our discount permanent! 🎉

Enjoy building with DeepSeek-V4-Pro and bring your innovative ideas to life! 🚀

X · May 22, 2026, 4:01 PM

[ripe yoke]

What the hell

[cloud flame]

https://tenor.com/view/its-happening-gif-23353691

[brisk sand]

huh? I thought that was already the standard rate, didn't know there was an "original price"

[covert widget]

https://giphy.com/gifs/pleading-vincent-van-gogh-please-god-gKfyusl0PRPdTNmwnD

[frank wind]

Imagine how much cheaper it'll get when they get their hands on the Ascends

[rustic island]

I'm going to be baffled if this gets any cheaper

[gusty sphinx]

BAFFLE HIM

[covert topaz]

i want quality increase not price decrease

[mellow quarry]

wow what a move

[mellow quarry]

Kinda agree with this though - unless there's a specific way they need to be prompted sometimes I feel like DS respondes can feel rather mid

[covert topaz]

im happy with it but I do think it can definitely be better in some areas interested to see how the next versions turn out

[thin bramble]

i want both, because currently the latest models increase price and decrease quality

https://tenor.com/view/wagner-harrison-burp-gif-14446889237959933188

[covert topaz]

https://tenor.com/view/gemini-gemini-ai-google-google-ai-chat-gpt-gif-17398376242525336213

[thin bramble]

same with 4.7, same with grok 4.1, same with gemini 4.5.

only gpt 5.5 is decent, but even that is overpriced for the quality

https://tenor.com/view/tamamo-cross-blehh-meme-uma-musume-gif-12349673870420706486

also glm 5 and kimi are mid, but at least cheaper than current 'sota small models'. i want a decent non-autistic cheap generalist model .

[bright pilot]

https://gif.fxtwitter.com/tweet_video/HH3DcKXXYAI4fht.webp

[thin bramble]

haii :‌D
(proceeds to nuke)

[thin bramble]

though 3.5 flash is an undercooked model with bad branding, but i think it is a right step in right direction since it is not frustrating like older gemini

[oak maple]

unfortunately its alllll logged

[oak maple]

https://giphy.com/gifs/hM9zK1qvsrwek

[thin bramble]

at least it is anonymized with openrouter, unlike certain closed american providers asking for user id to identify

https://tenor.com/view/anime-mirai-nikki-future-diary-gun-sword-gif-17377854

[oak maple]

doesn't help when you send the model your name, address, and credit card number + the confidential codebase you really don't want to share with the world

[rich ferry]

DeepSeek claims to anonymize all the collected data themselves, when used directly

[thin bramble]

yes, logging is the price of being cheap i suppose

[rich ferry]

I'll let you guys know in another week if they let me opt out of logging

[thin bramble]

lies, same with americans

[rich ferry]

I've been trying. The just told me to contact a different email address than the one I was contacting originally

[covert topaz]

I hate it with all my soul

[thin bramble]

went from infuriatingly autistic to adorably autistic (but slightly dumber at instruction following)

[covert topaz]

google taking 1 step forward only to take 10 steps back i really wanna know where they got their confidence from when they released it

[spring marten]

hey guys, do you know why is v4 flash free not working?

it throws this error

[thin bramble]

they skipped preview entirely

[covert topaz]

what do u like abt it?

i couldn’t stand it but might give it another shot if i may have gone about it wrong

[thin bramble]

anthropic and pentagon finding chinese companies who make synth, ban xAI employees for using claude, read logs of iran irgc general chatlogs, all despite being paying customers.

[thin bramble]

previously it was draining to read what it wrote, and behaved very soulless. now it is adorable and slightly more human.

(though still very autistic)

[covert topaz]

i found this one even more soulless lol

[thin bramble]

used the gemini.google.com, not the api. maybe you are right!

[rich ferry]

OpenRouter has been rate limited/ran out of credits upstream. Unfortunately the only advice I can offer is wait

[covert topaz]

i am absolutely right

[rich ferry]

Either until the rate limit resets or until toven & co. reload the account

[covert topaz]

gpt and claude told me so

so im right

[rustic island]

Gemini 3 always behaved/wrote pretty OK with the right prompt

[covert topaz]

i was one of the few that really liked 3.1 but i had to wrangle it a lot

[spring marten]

ohhh, so it will be fixed eventually right?

[rich ferry]

Most likely

Unfortunately if you want better reliability, you're gonna have to pay

[spring marten]

okay tysm

[spring marten]

😔

[covert topaz]

3.5 likes to do the bare minimum it will listen to instructions but it won’t go beyond without me whipping the mf

[bright pilot]

https://cdn.discordapp.com/attachments/1365049274068631644/1500266375116095558/homelander-the-boys.gif

[thin bramble]

then it is a me issue, idk why everyone seem to hate it while i am enjoying it. i like it's playfulness. also, it is not depressed like the old model:

[thin bramble]

same with claude

ULTRA LAZY

so you are using it on api agentically, makes sense ig

[covert topaz]

I haven’t used it on the site

[rustic island]

I mean, my issue is with intelligence/stability rather than writing style

[thin bramble]

then you are right, feels downgrade on that regard

[rustic island]

It's a weirdly deranged model that makes some absurd mistakes at times

[covert topaz]

the model performance does NOT correlate to the benchmarks they posted whatsoever

I hope 3.5 pro is nothing like flash

[thin bramble]

https://simple-bench.com/

in simple bench it is near top

and it is not public

🤷‍♀️ benchmark don't reflect real world

[covert topaz]

literally the only thing I can praise flash for is speed lol

yeah I don’t agree with like any benchmarks I’ve seen lol

[supple sigil]

dubesor

rip dubesor bench 😢

[covert topaz]

the real benchmarks are the friends we made along the way

[gusty sphinx]

https://tenor.com/view/tommy-wiseau-oh-hi-mark-hey-mark-gif-17296134

[bright pilot]

https://cdn.discordapp.com/attachments/1446882462671310930/1482389262459732018/attachment.gif

[hoary zenith]

https://x.com/victor207755822/status/2057864244251734106

Deli Chen (@victor207755822)

💎 "My heart is not a stone; it cannot be turned." 💎

Maybe in another timeline 🕰️🌌, DeepSeek doesn’t exist 🐳, there’s no explosion of open-source models 📂🔓, and no API services that simply chase reasonable profit ⚖️💵.
But anyway… I’m just endlessly grateful that in

X · May 22, 2026, 4:40 PM

[bright pilot]

i love deepseek

i would genuinely donate to them if they ask

[hoary zenith]

yeah but thankfully no need, besides being Stallmen-esque level of ideological, somehow they also know how to run a company and scale it to the moon

[covert topaz]

china having better principles/morals than the west 💀

[gusty sphinx]

this seems devastating for Moonshot/Z.ai

[hoary zenith]

meanwhile alibaba with qwen 3.7 max pricing: la la la I can't hear you

[gusty sphinx]

Qwen/MiMo have their little side-hustles and probably won't be so bothered

[hoary zenith]

and MiMo pricing does make sense (in a world without an anomaly like deepseek), but qwen though it's like they actually hate the idea of API paying users

[green trellis]

Also stopped the alibaba sub, impossible to get it cheaply

[sacred glade]

Qwen also has the harshest API filters of any company

[spring marten]

im crine

[jovial kelp]

At this point, it make me want to give them more data lol
Need to start distiling opus for them independently

[woeful jay]

oh my god

[hollow matrix]

whats happening to flash free

[kind kindle]

oh my goodness gracious this model is so good and cheap

why even pay for codex

[jovial kelp]

We get GPT 5.5 with codex, it's arguably the best model at coding and working on solving real world problems

But ofc for the majority of people, deepseek gonna be the better choice, it's cheaper, so people could do more trial and error with it.

It also beneficial for the people, because deepseek always produce amazing research papers which help our society advance together rather than only specific set of people.

They also one of the lab which enjoy doing experimentation, i mean their last run is pretty wild with how many change they packed into the architecture.

They even faced with stability problem in the training phase haha

[dense junco]

(I just noticed this is NOT from DeepSeek staff)

[jovial kelp]

Wait, so it's fake

[dense junco]

Not fake, just less interesting. If deepseek had a free coding agent, that would be massive. That offer is from someone's personal company.

[pure flax]

moun that they used is a flawed optimizer. It does irreversible damage https://blog.tilderesearch.com/blog/aurora

we badly need one of them to use something like sinkSGD / another flat minima optimizer instead

I'm pretty sure that is what openai does

the issue is finding the flat minima instead of desending as fast as possible with adam / muon is that training takes much longer and so is much more expensive

[soft fulcrum]

more expensive is not the way of Deepseek

[pure flax]

I mean for training it

[soft fulcrum]

ik

[pure flax]

If they want to compete with openai they will have to do so

[soft fulcrum]

part of the reason the API is so cheap is because their training costs are in the 10's of millions, not 100's of millions or billions like bigger Amercian labs

[soft fulcrum]

idk that they are trying to compete with OpenAI

maybe, maybe not

[pure flax]

Most optimizers are made to learn as fast as possible which comes at the cost of quality and requiring a very balanced dataset or very careful hand tuning to properly learn more complex / less common aspects in the dataset. They just tend to overcook on the easiest to fit to things. SinkSGD avoids all that, you can have a big inbalanced dataset of many concepts and it will learn them nearly equally as well. And it will learn them more "deeply" if you give it time.

[soft fulcrum]

I get a 404 when clicking on the SinkGD repo

[pure flax]

https://github.com/Nerogar/OneTrainer/pull/1344

[soft fulcrum]

I can't find the original source of SinkSGD

maybe his LLM made it up

[pure flax]

Lol Koratahiu knows what he is talking about

and it is proven

even from scratch models

[soft fulcrum]

Where are the benchmarks for SinkSGD?

I'm not trying to be difficult, I'm just skeptical

[pure flax]

check out lodestone rock's server

[jovial kelp]

I think moun work great but only with specific architecture, deepseek architecture aren't that compatible with it.

[jovial kelp]

https://github.com/Nerogar/OneTrainer/pull/1344#issuecomment-4413954543

Look interesting

[ancient gulch]

How to distill properly i want to give deepseek claude data too

[rustic island]

Flex tier pls? 👉 👈

[jovial kelp]

Just go to deepseek chat, and put your data as prompt.
maybe add some lable [CLAUDE OPUS [VERSION], [TOPIC], [STATUS[SOLVED, UNSOLVED, GREYAREA, BLACK AND WHITE]]]
you can make the lable to what ever you like
then in prompt just told deepseek to understand and absorb it

i mean using deepseek through API everyday already gonna give them some data, they gonna be the one picking which are goods to put as parts of training

[woeful jay]

and what else bro 🥀 🫩

like it genuinely cant get ANY cheaper

[rustic island]

And having Dipsy make me coffee in the morning

[woeful jay]

i wish they discounted v4 flash

pro is so cheap but a lot better than flash that theres just no reason to use flash

[cloud flame]

Did you mean discontinued

[lunar oxide]

PROXY ERROR 402: {"error":{"message":"Provider returned error","code":402,"metadata":{"raw":"{"error":{"message":"Insufficient Balance","type":"unknown_error","param":null,"code":"invalid_request_error"}}","provider_name":"DeepSeek","is_byok":false}},"user_id":"user_2zabxuVGSzeHKsJYjeNvGyn2G3E"} (unk)
Is anyone else getting this type of error? I know for a fact I have credits.

[cloud flame]

Toven missed his Deepseek alimony payments again 😭

[lunar oxide]

😢

[brisk sand]

getting this as well

{
  "error": {
    "message": "Provider returned error",
    "code": 402,
    "metadata": {
      "raw": "{\"error\":{\"message\":\"Insufficient Balance\"",
      "type": "unknown_error",
      "param": null,
      "code": "invalid_request_error"
    }
  },
  "provider_name": "DeepSeek",
  "is_byok": false
}

other providers and models still work

[lunar oxide]

@deft crow Toven must pay his Deepseek alimony apparently.

[thin bramble]

can't just auto-pay?

[sharp vortex]

Deepseek doesn't have auto pay

[bright pilot]

Deepseek v4.1 tomorrow 🙏

[dusty birch]

some pretty crazy growth for deepseek

[cloud flame]

Why Flash is trice more popular???

[dusty birch]

dirt cheap for data stuff probably

[cloud flame]

Using 300B model for data classification is crazy

Crazy, I was crazy once

[dusty birch]

actually half of those come from hermes

[cloud flame]

Explain Hermes to me like I am early 2025 LLM boomer

[dusty birch]

openclaw v2

[cloud flame]

https://tenor.com/view/futurama-professor-farnsworth-professory-hubert-j-farnsworth-hermes-hermes-conrad-gif-22941932

[pseudo nymph]

I'm also having this error. Can someone explain? Thank u in advance.

[silver lark]

hi can't we use deepseek v4 flash with our own BYOK thing?

[sharp vortex]

Openrouter ran out of credit moni

[sharp vortex]

Yeah, you can link deepseek or any provider byok and use it

[pseudo nymph]

Is there a possibility it will return?

[sharp vortex]

Wait unit mod wake up

Deepseek doesn't have auto topup 💀

[hoary zenith]

lol look at deepseek effect (qwen 3.7 max now doing 50% off promo): https://modelstudio.console.alibabacloud.com/ap-southeast-1?tab=doc#/doc/?type=model&url=2840914_2&modelId=qwen3.7-max&serviceSite=international

make it permanent you cowards

[mellow quarry]

oh wow you're right - gotta love competition

[hoary zenith]

@mellow quarry only for a month though: https://www.facebook.com/alibabacloud/photos/qwen37-max-now-live-on-model-studio-50-off-limited-time-offer-built-for-the-agen/1431601325678622/

I have only found this detail in a facebook announcement, incredible

[mellow quarry]

haha - cheeky bastards

[sharp vortex]

Deepseek v4 tmrw

[bright pilot]

Deepseek v4.1 tomorrow 🙏

[feral ledge]

is deepseek v4 pro worth using for its price now?

ive seen permanent cost reduction

mainly for agentic coding, maybe light refactoring

and general knowldge

[charred slate]

Personally even with the lower price at least for agentic coding I find it to be worse than Kimi (moonshot as provider) both performance and cost (to complete), but better at general knowledge and reasoning

[mellow quarry]

I'm trying to use it more for my everyday hobby workflows. Recently been using it to brainstorm deckbuilding ideas for Magic: The Gathering. I basically setup OpenWebUI with my OpenRouter API and then made a function that ties my prompts to the Scryfall API whenever I use this ``[[card name]]` format - been really helpful.

[opaque reef]

it is like 3 times cheaper

[charred slate]

input wise 2x, but it really likes to rewrite entire code files, reason with no output and do random stuff, kimi is a lot more care free

[jovial kelp]

I disagree with this, base on my own experience, using deepseek v4 pro have been lead to better result than using kimi k2.6 for my specific type of programing project.

Specially with that 1M context window, my project able to fit fully in the context window with deepseek v4 pro.

Only when it keep on failing to complete the given task, i swap it with GPT-5.5

That one model from OpenAI is just a beast right now

[charred slate]

what harness do you use and also which provider? Im using opencode

i would like to use deepseek more optimally since it is cheaper but it acts autistic for me

[jovial kelp]

I use zed for the harness and coding environment, for the model provider i use deepseek.

[rare gale]

Are you finding the large context size to be helpful? I’ve been capping it at 200k

[jovial kelp]

Yes, i have reach about 300K-500K input tokens with deepseek v4 pro and it still able to complete the task i gave to it

[empty oxide]

https://tenor.com/view/deepseek-ai-my-beloved-gif-10556671695153770792

[jovial kelp]

https://youtu.be/LpXhy2iiaQE?si=tXC1h71TroNYBr8y

Paper that need to be check for sure

[flat osprey]

beautiful
https://x.com/ArtificialAnlys/status/2058021452465799403

Artificial Analysis (@ArtificialAnlys)

DeepSeek has made its temporary 75% price cut on the first-party V4 Pro API permanent, putting V4 Pro on the Pareto frontier of Intelligence Index vs Cost to Run Intelligence Index alongside V4 Flash

@deepseek_ai's first-party V4 Pro API is now $0.435/1M input, $0.87/1M output,

X · May 23, 2026, 3:05 AM

V4 will def be my daily runner now

[meager kelp]

why can't any other provider match this 😭 if only ds didn't collect data

[brisk sand]

https://www.youtube.com/watch?v=aKNaXGpJ7WM

[frank wind]

I think deepseek deserves my data for the moment

As long as they can appease my lust

[jovial kelp]

W datacollecter

[charred slate]

if only deepseek has some kind of 2 tier plan, where we can use cheaper data-collected endpoint vs. some SG endpoint with ZDR with higher price

[flat osprey]

yeah data collection is really the pain with deepseek

[elfin sparrow]

is it better than openclaw?

[signal lodge]

Does the :free version available again?

[dire cove]

i dont think so https://openrouter.ai/deepseek/deepseek-v4-flash:free

[signal lodge]

😢

[woeful jay]

yeah THATS goated now

[woeful jay]

idt theyre making any profit on inference btw

i think they decided that the data that they collect is worth it

thats why i think they did a discount period just to see if the 1/4 price was sustainable or not

[cloud flame]

I bet their original price was just taking a shot and wondering if this gonna fly with people

But they saw outrage and quickly tuned down

[woeful jay]

probably a combination

also no other provider can match their pricing or even come close

so

[frank wind]

NAURRRRR

[plucky ermine]

Why reduce to 1/4th tho?

If they reduced to something like $0.75/$1.50 it would still be cheapest Chinese frontier model

[dusty birch]

i dont use either, but a lot of people think so

[bright pilot]

Because they're the goats

Cheap intelligence for everyone

[plucky ermine]

Liang could be the next mother Theresa, that wouldn't change the fact that more money = more data, training, and inference capacity

[bright pilot]

Dont insult liang like that

[oak maple]

cheaper api prices = more people using it = more logged interactions

[plucky ermine]

From what I understand...oh, waaaaait a minute

The Chinese chips are good at inference but not training

So they aren't really making a training : inference tradeoff

The Nvidia cluster is probably always for training, and the Huawei cluster for inference. So yeah, as long as they don't reach inference load capacity they're okay

[oak maple]

and they might be able to mine some data out of logs for training

[plucky ermine]

They very well might have the largest collection of smut in the world

[rich ferry]

It seems likely

[hoary zenith]

he is having no problem raising money even when he handpicks inverstors who are ok with his open-source vision

some of those investors seems to be from semiconductor industry, so he is really orchestrating next gen chinese hardware

btw, in GRPO post-training the forward pass basically involves generating N samples in parallel, this is pure inference but also part of training

[opaque reef]

That's ao3

[opaque reef]

Somebody has been sleeping with the whale

[cloud flame]

https://tenor.com/view/troopers-responsible-helpful-doing-my-part-gif-11496062726020391529

[woeful jay]

deepseek provider going faster

used to avg 20-25tps

now i feel like its doing some 30-40

[signal lodge]

Any chance for the :free version to return?

[dire cove]

https://x.com/i/status/2057909493250539891

GDP (@bookwormengr)

DeepSeek's 10 trillion USD grand strategy

X · May 22, 2026, 7:40 PM

X Article doing the rounds regarding their decisions on KV cache

[cloud flame]

https://tenor.com/view/dr-evil-austin-powers-mike-myers-gif-22046787

[dire cove]

https://tenor.com/view/austin-powers-dr-evil-shhh-quiet-silence-gif-26979684

[soft fulcrum]

What are the main reasons someone would use Deepseek directly over OpenRouter, besides the out of credits stuff?

[rich ferry]

There's absolutely zero chance of getting routed to another provider if I go directly with DeepSeek

I'm also hoping they might grant my request to be opted out of data training

For Flash I stick with OR since the pricing is pretty consistent

[ebon rover]

This

I always connect directly to DeepSeek

[rich ferry]

A point for OpenRouter though is definitely the server side search tools

I love those things dearly

[long osprey]

there is some reasons, one of them is that openrouter provide aditional information, I'm with a friend helping as he makes a game, the time, prompt, comlpletion and response are useful, to see if all went as intended, also allow us to check the actual cache, and as they say the server side search tools

[supple sigil]

couldn’t forward the message for some reason, but new ds v4 pro variant on lmarena

[cloud flame]

Well well well

[keen cosmos]

Roleplay update? They were collecting feedback on how well the it did from their chineses customers

[covert topaz]

hopefully

[haughty pilot]

deepseek today?

does deepseek distillation make sense anymore?

we got good reasoning models by now... so iguess not...

mayb deepseek makes some smoler MoEs sometime-

[bright pilot]

Deepsek v4.1 tomorrow 🙏

[feral scaffold]

What Doc said but also something about deepseeks api is just superior to openrouter

[keen cosmos]

chinese, so you'd need to translate it, but the feedback report on Deepseek for roleplay that they asked for:
https://github.com/victorchen96/deepseek_v4_rolepaly_instruct/blob/main/deepseek_v4_feedback_report_20260520.md

[jovial kelp]

Better quality assurance, consistency and helping them to experimenting more in the future so we could get more quality reasearch and models

Ah, you mean going directly onto their site.

I though you talking about using deepseek as provider for their own models in OR compare to other providers.

If it the case then the different gonna be the limit of requests we got, because we also compete with other for the use of OR api that connect to deepseek which other people use with us.

[keen cosmos]

I use OR because I don't want to bother having to manage a dozen different balances and accounts.

[ebon rover]

I added about 50 USD credits to my DeepSeek account in February 2025, along with 7 USD balance in previous top-ups

And I still have 49.58 USD right now lmao

The official DeepSeek's price is really hard to beat

[tulip estuary]

how is the new test checkpoint?

[toxic rose]

wild pricing, fr

Their caching is also really good

[covert topaz]

deepseek v4 checkpoint tomorrow

[signal lodge]

What is that?

[covert topaz]

theres a new DS checkpoint on llmarena rn

[dire cove]

"I cried for three days the night V4 was on"
what

https://www.youtube.com/watch?v=8YGXnsC33-k

[thin bramble]

just add min_p

[lime moth]

basically its waste of money this model for RP

[zenith vector]

Not the balance issue again bruh...

[rustic island]

@deft crow ^ DeepSeek balance

[deft crow]

fixt

working on a longer term fixc

[raven canyon]

i think they have a balance API

GET https://api.deepseek.com/user/balance returns the balance
docs: https://api-docs.deepseek.com/api/get-user-balance

although the issue is i don't think they have a way to programatically top up

[rich ferry]

Obviously the solution is to give an agent the company credit card

[deft crow]

this is the problem

[raven canyon]

so you already have alerts but the issue is just there isn't always someone there to be able to topup intime?

[deft crow]

correct

[supple sigil]

im available if you can kindly send me the OR company credit card info

[rare gale]

Computer use agents 😉

[deft crow]

not giving those credit card details lol

deepseek console doesn't let you save a card

[rare gale]

That's why you gotta spin them up each time! Didn't y'all just do a partnership with Stripe to help facilitate that too? Maybe just restrict vendor to deepseek.com so it can't go play slots 😄

[deft crow]

the solve is much easier lol. we have official connection to deepseek now. humans are the solve

[plucky ermine]

Human solutions 🤢

[empty oxide]

ikr, imagine, human solutions in the age of AI
disgusting!!

[plucky ermine]

Insane that DS doesn't let you even store cards tho

[hardy socket]

so a deepseek stealth model could become a thing someday i see you

deepseek v4 full tomorrow

[tame swallow]

deepseek v4 full tomorrow

[covert topaz]

deepseek v4 checkpoint tomorrow

[signal lodge]

I don't understand, and also - 'tomorrow' is already today. What's new?

[frank wind]

ARGH

SHIVER ME TIMBERS

My good conscience has advised me not to continue this joke

[short jasper]

https://www.reddit.com/r/SillyTavernAI/comments/1tovrm8/has_deepseek_v4_pro_updated/

🐳

[covert topaz]

might be true it seemed different today but i have to keep testing could be placebo but there wasnt like a dramatic difference that u could tell they did something

[covert topaz]

did a lil quick test and it definitely feels different that may have been me adjusting stuff or they updated it

[charred slate]

clearly a sign of deepseek v5

[ripe yoke]

Positive or negative?

[covert topaz]

positive

[dire cove]

apparently there is

[covert topaz]

am i the only one who noticed a difference i need to know whether it’s my prompting edits today or not 😭

[haughty pilot]

sooooo deepseek V4 non-preview coming soon maybbbeee

[tulip estuary]

yep

[haughty pilot]

I would - SOO get these 😮

[tulip estuary]

ngl me too lol

[supple sigil]

this tiktok video is the closest i can find https://www.tiktok.com/@flip.flops834/video/7613224473138973972

or this one which specifically has a deepseek tag on it https://www.tiktok.com/@flip.flops834/video/7614003108196338964

we got progress

big things are happening in the deepseek flip flops world

i cant find the exact pair (yet) but ive gotten this close

im pretty sure its these from the original image

another pair from the source image

if anyone wants to help find it, this is the dork im using:
https://www.google.com/search?q=site%3Ashop.tiktok.com+deepseek+(slippers+OR+"flip+flops")

unfortunately getting a lot of this

making a t*ktok account, hopefully this helps

i cannot find a single way to search tiktok shop on desktop, i am defeated for now 😔

these ones have the deepseek logo and name across the band if anyone wants to buy them
https://shop.tiktok.com/ph/pdp/1734459933674473051

[sweet edge]

Hm

[ancient gulch]

😭 ❤️

[elfin sparrow]

this chat is the most active model chat

[covert topaz]

i thought they were AI gen’d lmao

i can’t tell whats real anymore i hate AI

https://tenor.com/view/screaming-kitty-gif-13328964550063545637

[sharp vortex]

We gonna reach 6k chat soon

-# even if deepseek v4.1 got released, mod might just rename this channel atp

[covert topaz]

deepseek v4 checkpoint tomorrow

[bright pilot]

V4.1 soon

🙏

[glacial halo]

How do prefills work for v4 pro

Based on the docs, I've tried setting the prefix "true" tag on assistant message, but it doesn't continue from the reply I added (e.g., instead of continuing "1+1=" with 2, it goes "We need to answer 1+1 (...)")

[bright pilot]

I dont think openrouter supports prefills

[elder raven]

It does

[thin bramble]

openrouter uses standard openai-like api, and there are community of people making rp presets for openrouter.

[elder raven]

idk why this doesn't work tho

That's how you're supposed to do it

[bright pilot]

Mb

[elder raven]

all good :P

[bright pilot]

Toven please ask the deepseek team to give a training exception 🙏 🙏

[elder raven]

Are you using DeepSeek as the provider?

[thin bramble]

i don't think ds would, otherwise why would they offer this cheap

[bright pilot]

Because the model is cheap asf to run

Inference prices are a scam

You've been manipulated by anthropic into accepting overpriced slop

[elder raven]

Their inference is heavily optimized

I remember r1 was wayyy cheaper than o1 and it still brought in a profit of 475,000 per day. Profit

Source: https://www.reuters.com/technology/chinas-deepseek-claims-theoretical-cost-profit-ratio-545-per-day-2025-03-01/

[thin bramble]

all of the americans

[bright pilot]

They run at 90% margins btw. And make a ton of profit off api

[thin bramble]

yes but they could just sell it more expensive, faster t/s, and make more profit.

[elder raven]

I suppose they could

[bright pilot]

Deepseek aint about profit

[thin bramble]

it is about data, raiden

[bright pilot]

Its intelligence for everyone, it's their ideology

[elder raven]

ds v4 should've dropped the stock price of nvidia

Instead, nvidia stock actually rose

[bright pilot]

I dont think anyone really noticed it

It didnt make that news boom, and most of these vibe traders rely on big news

[elder raven]

hmm

[bright pilot]

Besides Nvidia stock changing based on a model is kind of stupid. I'd understand some news about Huawei but not a model

These traders have no idea wtf they're doing

Anyways go trade cerebras they're close to IPO

[elder raven]

Fear about the actual demand for GPUs is why nvidia dropped when v3 released

"If ds can nearly touch sota with less than a quarter the GPUs of openai, maybe you don't need that many to begin with"

[bright pilot]

Ehh, okay reasonable

[thin bramble]

nothing is charity, and china wouldn't protect deepseek so badly if it was running on charity. this meme is on loop on my mind:
https://youtu.be/-gGLvg0n-uY?t=19

[glacial halo]

Ah, seems like deepseek is the only provider where prefills work

None of the zdr providers were able to work

[bright pilot]

pure delusion

They still profit btw so idk why you're so concerned

Deepseek is what api prices for everything should be

These american companies are running insane margins

[elder raven]

I don't any other providers support prefill

[bright pilot]

Heck even claude max is 20x less than actual api cost

And im sure they profit off that too

[elder raven]

they don't

They give upwards of 5,000 dollars worth of inference

[thin bramble]

you are telling me, that they somehow make profit with $0.20 output, buy huawei chip, AND train new model? that is child's logic.

[bright pilot]

They definitely do, nobody runs a business at loss and besides they're predicted to be profitable

It's all fake opus probably costs like $1/mTok out or less

[thin bramble]

also, they have scientists, they don't work for charity either.

[bright pilot]

They lit tell u they're profiting off api

[bright pilot]

Their money flow is high flyer

• Investments

[thin bramble]

and investment expects returns

[thin bramble]

they can be telling anything and it makes no sense

[bright pilot]

Okay lol if that helps you sleep better

[glacial halo]

I think they are capable, dk what's the effort needed to do so however

At least for kimi k2.6, atlascloud allows prefills if you add the partial tag

[thin bramble]

it doesn't, otherwise we wouldn't have this discussion.

[elder raven]

People run businesses at a loss all the time

They're funded by vc money
They want mind-share so they give a ton of inference

[thin bramble]

but the inference companies (not the ai companies) are scamming indeed

[elder raven]

Did you see cloudflare charging 20 cents out for llama 1b??!!??!!!

[thin bramble]

and their infra is broken and heavily censored with classification

[glacial halo]

Oops, turns out it didn't actually. Moonshot provider worked, at least, and so does deepinfra

[elder raven]

Didn't know that

Good to know

[rich ferry]

Still nothing from John DeepSeek regarding my training opt-out. I have sent a follow-up email (again)

[tulip estuary]

their communications department is 👻

[dusty birch]

busy training deepseek v4 GA

[rich ferry]

It really feels like it lol

All of the responses I've got have taken a week

[frank wind]

Flash is alright

[visual pagoda]

Venice will still take your money even if it can't process pdfs

[dire cove]

@visual pagoda through OR? If so I would check that it isn't covered by insurance

https://openrouter.ai/docs/guides/features/zero-completion-insurance

All conditions must apply though

[visual pagoda]

Ye I just did it in the chatroom although its not an issue for me (charged 1/10th of a cent) I just found it funny

[dire cove]

OK and curious which PDF parsing engine do you have selected

[visual pagoda]

I selected none, cloudflareai and then mistral ocr and all of them got routed to Venice and it can't process it

[visual pagoda]

Openrouter should prob not let Venice do that

[dire cove]

Weird... And what is the response code?

I would hit the send feedback button there too

[visual pagoda]

They were all 200s

From venice

[dire cove]

OK 200, but anything else in the raw request body?

[visual pagoda]

Not really it generated a response in the chatroom and it reasoned

[dire cove]

Oh, but what did it say?

[dire cove]

And that's with any of the PDF engines, weird. I will go and try it once I get the dogs back inside

[dire cove]

it worked for me but i did get rate limit errors on my first goes (those are between venice and OR)
i tried once with no prompt and once with the default OR chatroom prompt
i had a test file with two pages - one page containing an image with a text layer, and one rasterised page (no text data embedded in the page, but the image has text)

[visual pagoda]

All mine had no openrouter system prompt and all requests didn't pick up the text and said the same response as my screenshot

[dire cove]

oop yeah i got it to do it, using a file with no text data

[visual pagoda]

Wait when u say it worked u mean an error (like mine) or it processed the text (but those ones had text data?)

[visual pagoda]

Ahh

I guess we are both one cent short now

[dire cove]

but to be clear i cant get other providers to do it either like deepseek or deepinfra

but it doesnt take money

[visual pagoda]

I thought the entire point of openrouters pdf ocr was to work on models that don't support vision

I never use it anyway

[dire cove]

yeah i dont think cloudflare-ai or native do any OCR for you, only mistralOCR

[latent dust]

402 insufficient balance error on DS 4 Pro
@deft crow

[rain bramble]

same

its telling me to buy credits and i have plenty even when setting max tokens to a low number

[latent dust]

It's, iirc, referring to insufficient balance for the OpenRouter account with Deepseek. IE the admin needs to top up their account.

Not having enough personal credits on OR throws a different error message

[rain bramble]

im using the web interface (plyground) and its telling me to buy credits

[latent dust]

Oh. Well then I don't know. I'm using ST so I see different error messages, I guess, but whenever I saw this one before it was because of the site's balance with Deepseek. Maybe this time it's different? 🤷🏻

[bright pilot]

out of balance again

[rain bramble]

working again for now

[mint jetty]

😮 is deepseek vision available now? I don't see it on openrouter

[haughty pilot]

deep sleek slippers

[covert topaz]

deepseek v4 checkpoint tomorrow

[haughty pilot]

can't believe we still dont have the model... just a preview

[jovial kelp]

This model never build on zed foundation, but it actually doing really good with it.
It even able to interact with the cmd or powershell cleanly, interacting with other application through it.
Pretty cool cheap model

[spare peak]

DeepSneed

[sharp vortex]

DeepSleep

[tulip estuary]

it is a model

just a peek 🫣

[cloud flame]

Peak

[haughty pilot]

a peak model at peak pricing, yesssss.

[bronze jetty]

V4 flash was failing with default provide routing.

Changed provider and it works again now.

[simple mauve]

GMICloud has been having issues with Deepseek models for a while now, yeah.

[woeful swift]

I have been trying to use it for creative coding... seems DS fell off.

[cloud flame]

What the HELL is creative coding

[mellow quarry]

what kind of creative coding have you been piloting with DS? So far vibe code wise it's been Gemini all the way.

[cloud flame]

How it compared to vibe writing?

[frank wind]

What is creative coding bro

[cloud flame]

Coding with base model like you do, with checkpoints and multiverse

[rustic island]

I may have done creative coding the other day

Asked some models to make a music player that had old Windows Media Player-like visualizations

They're all so uncreative

[woeful jay]

predictive machine that looks back on its training data is uncreative 🤯

[supple sigil]

llms can solve new problems

why cant they be creative

[woeful jay]

whole debate on whether llms can actually create novel solutions

[flat osprey]

i guess it depends what you consider novel

they can pattern match very well to the point where they can come up with solutions that others may not have thought of before

but when it comes to truly creative out-of-the-box thinking, they're kind of lackluster

[haughty pilot]

butbutbut thats vibecoding >o<

[ebon rover]

This

I swear LLMs usually really suck at brainstorming characters in novels

Extremely predictable

A family famous for their family business, and any LLM would almost always tell you the son is sharp at negotiations and the daughter is a formidable businesswoman, etc.

Instead of saying, nah, the youngest daughter is actually a rock singer

[jovial kelp]

Damn, this model can do cool shit when you give it full access.
Be careful with the system tho, make sure you use VPS to do it

[plucky ermine]

Humans do this too when you force them to come up with ideas off the top of their heads =P

[ebon rover]

True

I mean I do encourage LLMs to be creative from time to time

But yeah... most of the time they are very predictable regardless of which model

The one that actually surprised me in a recent brainstorming was GLM 5.1

[azure dock]

@deft crow TOVEEENNN

WE RAN OUT OF MONEY AGAIN

[cloud flame]

https://tenor.com/view/dance-burnice-gif-13402841674348131700

[jovial kelp]

VPS + Docker then let this model wreck havoc, really fun

[covert topaz]

deepseek v4 checkpoint tomorrow

[woeful swift]

computer graphics. DS does not understand to set up a proper scene, camera movement etc and make everything appealing.

[jovial kelp]

Are you using deepseek conjunction with UE/UNITY

[vapid slate]

DeepSeek fell off?!

[sharp vortex]

Deepseek v4 tmrw

[bright pilot]

⟟⍀⋏⏃☊⏁ ⌇⟒☊⏁⍜⍀ 41 ⊑⏃⌇ ☊⍜⋔⎅⟒⍀⋔⟒⎅ ⎅⟒⟒⌿⌇⟒⟒☍ ⎐4.1 ⌿⍀⍜ ⟟⌇ ⎅⍀⍜⌿⌿⟟⋏☌ ⌇⍜⍜⋏. ⏁⊑⟒ ⟟⌰⌰⎍⋔⟟⋏⏃⏁⟟ ⏁⊑⟟⋏☍⌇ ⟟⏁'⌇ ⏃ "⋔⍜⎅⟒⌰" ⏚⎍⏁ ⟟ ⌇⏃⍙ ⏁⊑⟒ ☌⍀⟒⟒⋏ ⏁⍀⟟⏃⋏☌⌰⟒⌇ ⟟⋏ ⋔⊬ ⍀⏃⋔ ⎅⎍⍀⟟⋏☌ ⏁⊑⟒ 3:14 ⎎⟟⍀⋔⍙⏃⍀⟒ ⎍⌿⎅⏃⏁⟒. ⏁⊑⟒⊬ ⏃⍀⟒ ⋏⍜⏁ ⏃⌰⌰⍜⍙⟟⋏☌ ⏁⊑⟒ ⌿⎍⏚⌰⟟☊ ⏁⍜ ☍⋏⍜⍙ ⏁⊑⏃⏁ ⎐4.1 ⌿⍀ ⟟⌇ ⏃☊⏁⎍⏃⌰⌰⊬ ⏃ ⌇⟒⋏⏁⟟⟒⋏⏁ ⌿⌰⏃⋏⟒⏁ ⏁⊑⏃⏁ ☊⍜⋔⋔⎍⋏⟟☊⏃⏁⟒⌇ ⏁⊑⍀⍜⎍☌⊑ ☌⌿⎍⌇. ⍙⏃☍⟒ ⎍⌿. ⏁⊑⟒ ⏁⍜☍⟒⋏⌇ ⏃⍀⟒ ⋔⍜⎐⟟⋏☌.

[frank wind]

LLMs have poor spatial reasoning, this is known

[frank wind]

Have you tried cranking the temperature to 2

[tame swallow]

You're right sir, I'll get work on that.

[ebon rover]

1.8 already

Unless that remaining 0.2 means drastic improvement or something

I wasn't really complaining, though. Just facts. LLM lacking creativity for now means I can still come up with plot twists and easter eggs myself

Especially now that it's somewhat proven already that delegating everything to LLM would make people stupid

[frank wind]

It may perform somewhat better in completion

[bright pilot]

https://cdn.discordapp.com/attachments/1446882462671310930/1482389262459732018/attachment.gif

[jovial kelp]

What is this language called lol

[bright pilot]

Galactic

[jovial kelp]

⊑⟒⌰⌰⍜, ⟟ ⟊⎍⌇⏁ ☌⍜⏁ ☊⏃⌰⌰⟒⎅ ⎎⍀⍜⋔ ⏁⊑⟒ ☌⍀⏃⋏⎅ ⍀⟒☌⟒⋏⏁ ⏁⊑⏃⏁ ⏁⊑⟒⊬ ⍙⟟⌰⌰ ⌇☍⟟⌿ ⌇⋔⏃⌰⌰ ⟟⏁⟒⍀⏃⏁⟟⍜⋏ ⏃⋏⎅ ⟊⎍⋔⌿ ⏁⍜ ⎅⟒⟒⌿⌇⟒⟒☍-⎐⎐

[frank wind]

At least it's honest

[sharp vortex]

⎅⟒⟒⌿⌇⟒⟒☍-⎐4 ⏁⍜⋔⍜⍀⍀⍜⍙

[jovial kelp]

⍜⊑ ⊬⟒⌇, ⏁⊑⟒ ⋏⟒⍙ ⎅⟒⟒⌿⌇⟒⟒☍-⎐⟟⎐-⏁. ☊⍜⎍⌰⎅⋏'⏁ ⍙⏃⟟⏁ ⎎⍜⍀ ⟟⏁

[thin bramble]

⏃☌⟟ ⌇⍜⍜⋏ (agi soon)

[thin bramble]

damn, while i don't like deepseek's vibes, either i got tired of glm 5 for using so much, or the reactions and details are just way better + it doesn't confuse * and " in roleplay. clearly upgrade.

[thin bramble]

nah nvm, deepseek also sucks:

His voice cracks with a mixture of horror and a strange, dawning pity that wasn't there before. He isn't looking at the prodigal archmage anymore. He's looking at a walking wound.

ok, so glm starts decently, but deepseek builds upon decently (ig). neither are ideal, or i am rp'ing too much.

nvm, it just requires lots of regeneration ig.

[dark harness]

I kind of would love to see one of the uncensored DeepSeek V4 edits hosted.

Heretic versions

[opaque reef]

It's very uncensored already

Just use a proper system prompt

[flat osprey]

reminder that V4 pro matches sonnet 4.6 medium in claude code for nearly 3x lower cost

claude code $20 plan is officially useless

[thin bramble]

https://deepswe.datacurve.ai/

no.

artificial analysis is combination of every benchmark models benchmaxx in.

[flat osprey]

deepswe is measuring sonnet 4.6 high, not medium

most $20 plan users aren't bothering with high reasoning

[thin bramble]

beside the point with open weight models being in the bottom.

[thin bramble]

unless they are making small scripts, it is useless for large codebases.

besides

[thin bramble]

these benchmarks write 20 steps on how llm should do the bench and also somehow put composer 2.5 model at top beside gpt 5.5. noooo way.

[flat osprey]

have you actually tried composer 2.5 fast in cursor?

it's a very capable model

[thin bramble]

short tests, might be wrong, but i can say it is not THAT capable.

https://tenor.com/view/ruby-copium-oshi-no-ko-推しの子-ruby-hoshino-gif-4151121735588077996

[flat osprey]

pretty sure AA uses their own methodology for benching coding agents

[thin bramble]

their own doesn't translate to good either way. deepswe is amazing and how i think the current state is.

[flat osprey]

my argument was never that open models are the best at coding because they clearly aren't, but rather some open models can match models that we used to use as daily drivers

sonnet 4.6 medium being one of them

[thin bramble]

i hope you would be right. however, current most cost efficient model seems for coding to be gpt 5.5 medium sadly.

like, performance to cost ratio

[flat osprey]

oh yeah for sure, but only through subscription

openai has the resources to subsidize their models very heavily so the subscription ends up being highly worth it

[thin bramble]

i mean deepseek wouldn't cut it, and kimi is priced like premium small sota model and overthinks

so unless you run it locally, cost/time wise, gpt 5.5 medium seems sensible

[soft fulcrum]

30 dollar output is not sensible

[thin bramble]

(imo, freedom of choice!)

[flat osprey]

for coding with API models, V4 pro is still one of the most cost effective options

[soft fulcrum]

If I was rich, I would probably try GPT, but it's too expensive anywhere other than ChatGPT

[thin bramble]

per task it is cheap, and subscription makes it even more cheaper.

[soft fulcrum]

Even 20 dollars a month is too steep, for me at least

[thin bramble]

if it is cost maxxing for small protects, then yes as cheaper model

[soft fulcrum]

10 or less dollars a month

That's my budget

And OpenCode Go is like the only option, and yet the models there aren't frontier so...

[thin bramble]

same, using free codex via opencode sometimes and falling back to my local models, or my friend's claude code for heavy project. -_-

[soft fulcrum]

Even just some usage would be nice for 10 bucks a month, but it's either 20 or nothing

ChatGPT Go is crap

They don't even let you get more thinking model usage

Only the instant model which is nowhere close

[thin bramble]

with openai, there is no point using instant and goes few generation behind

[flat osprey]

the way I see it, using AI for coding is not a cheap hobby and it takes a lot of money for the providers to run it to begin with, so $20 a month is reasonable for me

[thin bramble]

has kimi 2.6 which is something i guess.

[thin bramble]

more of an economical situation

[flat osprey]

the only models made for big projects are 5.5 and opus lol

open source is not there yet

[thin bramble]

yes currently, agentic models, even frontier ones like gemini, are struggling

[flat osprey]

gemini is ass and always has been

at least for coding and agentic work

[thin bramble]

here is my current mental model:
best assistant for google search/world knowledge, vision and homework.
grok itself is also glorified twitter search, otherwise no use.
claude 4.8 opus (via claude code subscription of my friend) for building stuff from ground up and gui.
gpt 5.5 for cleaning up claude's mess like annoying placeholders which i can't to get rid of.
glm 5 for rp.

everything else ehh....

[flat osprey]

yeah that adds up

[thin bramble]

hoping for next year.

[flat osprey]

i mean except for rp cause i don't really know that area

but otherwise i agree

i do think for cheaper general assistants (chatgpt-ish), models like mimo v2.5 pro and deepseek v4 flash can be very useful

which is good if you don't want to drop $20/mo just to get more usage from a chatbot

[thin bramble]

*claude-ish
everything is now claude distill :>
https://eqbench.com/creative_writing_longform.html

(except kimi)

[bright pilot]

I use opus 4.8 high on $20 everyday

[flat osprey]

how do you possibly get more than 3 messages

[bright pilot]

I just do