Qwen3-Coder-480B-A35B-Instruct

[iron estuary]

New Qwen coding & agent specific model with 1M token context length, now available on https://chat.qwen.ai and https://app.hyperbolic.ai/models/qwen3-coder-480b-a35b-instruct.

My initial testing indicate performance close to Kimi-K2, impressive for 1/2 the total parameters.

[iron estuary]

I used Qwen3-Coder-480B-A35B-Instruct to generate a procedural 3D planet preview and editor.

Very strong results! Comparable to Kimi-K2-Instruct, maybe a tad bit behind, but still impressive for under 50% the parameter count.

Creds The Feature Crew for the original idea.

[mossy pawn]

oh shoot

[hoary prairie]

impressive for 1/2 the active parameters
it has 1/2 the total parameters and 3b more active parameters

[winter shore]

bro

[hoary prairie]

my condolences

officially released btw https://x.com/Alibaba_Qwen/status/1947766835023335516

Qwen (@Alibaba_Qwen)

Qwen3-Coder is here! ✅

We’re releasing Qwen3-Coder-480B-A35B-Instruct, our most powerful open agentic code model to date. This 480B-parameter Mixture-of-Experts model (35B active) natively supports 256K context and scales to 1M context with extrapolation. It achieves

X · Jul 22, 2025, 9:12 PM

[torpid edge]

Don't care, give me the official API, I am feeling like just donating them money 😛

I guess tiered pricing is all the rage in China, I think bytedance doubao does this shit too

Must suck to implement this for you though

[winter shore]

yeah i was ready to take it live when they announced… i should’ve asked about pricing way earlier

i was not expecting 4!! tiers

we can do the two tiers now since gemini etc do it..

[magic linden]

oof

significantly more expensive than Gemini 2.5 Pro at full context....

[cedar remnant]

is dis model on some sort of api anywher

[torpid edge]

Hyperbolic

[cedar remnant]

is it on openrouter yet

[winter shore]

soon

[torpid edge]

hmm Hyperbolic doesn't even have prices for this yet lol

[cedar remnant]

okay

[brisk glacier]

nah.... this ain't it chief

[torpid edge]

tbh tiered pricing is closer to reflecting their actual computational cost

[covert schooner]

thats like double deepseek's, while being a smaller model
lets see what other providers will cook up

[silk flame]

They're probably pricing it based on the benchmarks. It's supposed to be just a little behind Claude Sonnet 4, but better than Gemini or GPT 4.1.

[torpid edge]

Good chance we may see some sub 0.5/M for input

[covert schooner]

yeah, seems unlikely to that for up to 128k, prices from other providers wont be competitive with deepseek

higher than that, probably not

[torpid edge]

Problem is Claude does implicit context caching which becomes even cheaper at 0.3/M

With open-weight agentic models catching up, the ball is in the court for some of the providers to actually start implementing context catching

[covert schooner]

did any even implement one like claude's?

because if no, they'll seriously have to work on that

[torpid edge]

No, I mean DeepSeek and Kimi does, but none of these Western inference providers to my knowledge

[trail glen]

what is qwen 3 coder plus?

[fallow nest]

it has like 10x the context length so its surely worth it

[snow coyote]

whyyy

[snow coyote]

can we get better prices with other providers?

[hoary prairie]

Most likely yes

[fallow nest]

A bit and we’re at Opus 4 prices

[winter shore]

https://openrouter.ai/qwen/qwen3-coder

[torpid edge]

Btw, recommended params:

temperature=0.7, top_p=0.8, top_k=20, repetition_penalty=1.05

[gaunt ore]

will this be implemented as the default through openrouter?

[open lotus]

qwen3 is really killing it now

newest coder is amazing

[mossy pawn]

you tried qwen code?

[open lotus]

its on hyperbolic

[mossy pawn]

i mean cli

qwen code cli

[open lotus]

ah no

[snow coyote]

is it really in same quality as claude 4?

[open lotus]

its really good

though they sure know it putting the price like it is in comparison to deeepseek

1 / 5

Hopefully another provider offers it at a better price, it would be cheaper to run that deepseek

[neat jasper]

this model so benchmaxxed 😪

[open lotus]

? my usual tests worked really well with it

[snow coyote]

i see hyperbolic is cheap

[snow coyote]

which tests you did?

[open lotus]

code of my own from some stuff that all but gemini 2.5 pro failed at

[soft birch]

They say it outperforms kimi k2 in benchmarks but I personally don't experience that at all

[hoary prairie]

not yet unfortunately, but they are looking into this in the future

[soft birch]

It agrees way easier and makes more often than not broken code that requires additional tokens to be fixed 💀

[torpid edge]

Really? I was quoting from the model's HF page: https://huggingface.co/Qwen/Qwen3-Coder-480B-A35B-Instruct

[hoary prairie]

Yes and that is correct just openrouter will not set it by defualt yet

I really hope they add it soon

[torpid edge]

oh right yeah

[hoary prairie]

oh replied to wrong msg

sry

[torpid edge]

Unfortunately at this point I have basically implemented a lot of openrouter myself somehow

[winter shore]

hmm? what parts?

[hoary prairie]

Setting recommended default parameters as default, instead of temp 1 etc

[torpid edge]

Automatic sampling parameter adding, the annoying /think and /no_think stuff for qwen 3 etc

[hoary prairie]

Good thing that they are leaving the /think stuff behind

Much simpler

[soft birch]

agreed

[snow coyote]

is it worth using it instead of copilot(claude 4, gpt 4.1)?

with cline

[hoary prairie]

too early to say for sure

[soft birch]

personally id say kimi k2 is worth using it more than this still

[snow coyote]

idk if 64k is enough for agentic or not

but price is good

[soft birch]

price for both is amazing compared to claude opus 😂

not referring to highest tier of qwen3

[torpid edge]

I hope DeepSeek v4 chills out and takes a month or two more, I am fatigued

[snow coyote]

yeah unbelievable at this price

[soft birch]

also btw with kilo code I never exceed above 40k per task

and usually agentic can condense context

[snow coyote]

I'm going to try thx for kilo

[neat jasper]

This

[snow coyote]

is there way to choose specific provider in kilo ?

[soft birch]

Yes, if you expand the "Advanced settings" at the bottom, you can select the provider

it will then update accordingly whether that provider supports prompt caching or not etc.

[snow coyote]

nice

this is so cool

[brittle adder]

How does Qwen3 stack against Opus?

[soft birch]

worse in quality, better in price

kimi k2 better imo

[torpid folio]

Adam Holter's benchmark

[soft birch]

Really? Qwen 3 is better than kimi k2?

[torpid folio]

On this benchmark, yes

[soft birch]

surprising

[torpid folio]

(This is mostly a coding benchmark)

[solar token]

add. caching. now.

[silk flame]

Do you have a link to the benchmark?

Unsloth quants are coming out: https://huggingface.co/collections/unsloth/qwen3-coder-687ff47700270447e02c987d

[torpid folio]

https://docs.google.com/spreadsheets/d/1FKO1i63BO-8353_wP7iAlowrSxdcT3ZAIZZbzARuybo/edit?usp=sharing

[torpid edge]

They haven't really managed to purge out the thinking tendencies:

evaluate_expr_tail([], Acc, Acc).
evaluate_expr_tail([add|Ops], Acc, Result) :-
    term(Term, _), % This is a bit tricky - we need to extract the next term
    % Let me rewrite this more cleanly
    fail.

% Let me rewrite the parser to make evaluation easier
% Simpler approach: build binary tree directly

[solar token]

Qwen coder seems like a model that clears kimi k2 and matches sonnet 4 - I left a showcase of what I built with it in #app-showcase 🐐

[glad reef]

How is it at non coding tasks

[dull sundial]

Qwen3-Coder is available in multiple sizes
I suggest using a more specific model slug other than qwen3-coder since other sizes will follow.
qwen3-coder-480b-a35b would allow for other sized qwen3-coder slugs

[somber spindle]

what’s the point if the tiered pricing is as it is right now

[winter shore]

yeah I flagged this internally

once we get the other models will probably make the change

you can call the model rn with qwen/qwen3-coder-480b-a35b-07-25

[hushed spear]

I tried the first prompt which is labeled as fail for qwen3-coder and this is what I get at one-shot

https://chat.qwen.ai/s/deploy/8a0db5e0-9fb0-4f5a-9d12-6d9c5a2c0d86

Some small issues like I can't select a tower and upgrade or sell it. But it's definately playable

[boreal pasture]

I think it did better than kimi k2 at my chess game prompt

Used hyperbolic provider with unsloth recommended settings

[hushed spear]

It seems currently all 3rdparty providers are all using fp8, while the model itself is bf16 originally.

[astral ingot]

running inference in fp16 is expensive and diminishing returns in terms of performance

[iron estuary]

I'm using the model in bf16 from the 1st party provider.

Incredibly strong showing creating a web OS! Probably the best result I have ever seen, period.

Notice the window reordering, resizing and minimizing logic at the end, it’s the first model to ever get this right. Also allows multiple instances of a given app with separate state.

This was created one-shot using the prompt:

Using Python, create a website that emulates an operating system.

Upon opening this website, a user should be able to view a desktop environment and use simple but WORKING applications such as a file browser, a text editor (which can edit and save text files), a web browser, a terminal and a calculator.

All of these applications should have at least all basic functionality working.

The operating system should have GUI elements, and have the general asthetic and feel of an OS.

Implement all the features listed above.

[astral ingot]

impressive one shot, probably one of the best ive personally seen

[boreal pasture]

Alibaba provider somehow did worse on my test than hyperbolic

Ill try again

Better now

Rather inconsistent

[solar token]

@winter shore can we expect caching for qwen 3 coder to be supported on Open Router?

[inner patio]

Don't mind me im just posting so the channel stays in my list 😄

[dull sundial]

Tested Qwen3-Coder-480B-A35B:

As expected from a coding focused model - most concise Qwen3 model

• 46 % less tokens than DeepSeek V3 0324
  While competent for general use, too, performed best in STEM (math) and coding obviously.

During creation of demo pages and further probing, it showcased several obvious weaknesses such as producing buggy collision, glaring UI oversights in multiple projects, and in general required error correction that was not necessary on models such as DeepSeek V3 0324.

For a massive, coding specialized model I personally was not convinced by its coding results, combined with the quite poor price/performance on current API offerings.

However, as always - YMMV!

[neat jasper]

You’re always spot on, thank you mr Dubesor

[rich field]

It's also on Chutes now

[tight bane]

For those providers using vLLM, the best way to do caching is with LMCache.

[atomic rune]

You might want to start running your tests with a specific provider / bench date and noting that in the benchmark, given how we've seen a lot of inconsistency in providers lately...

(assuming you're not running it yourself)

[formal elk]

does this model have a cache? if not, if this is not > sonnet and 2.5 pro, its Overprice

[solar token]

It does but openrouter doesnt support it, I've been asking since yesterday if they plan to add it

[atomic rune]

openrouter doesn't have or does not have a cache

it's the providers

if the provider you're using supports caching, then openrouter will support it too

[solar token]

https://www.alibabacloud.com/help/en/model-studio/context-cache#:~:text=When using text generation models,period will be periodically cleared.

[atomic rune]

should work then generally unless there's a specific new api they're using

or if they told you it's not supported yet

might be the weird pricing model that's making it hard

[solar token]

I think open router just doesn't support it, it's the same for kimi k2, they still don't support it

[atomic rune]

gotcha

[winter shore]

no we do need to implement it

[atomic rune]

ah ok

[winter shore]

pick the caching token values from the upstream usage and use it correctly + add the pricing

haven't gotten to it yet

for existing providers it's just do the thing

but for moonshot no

lemme look at alibaba tho

[atomic rune]

got it

huh, that should really be something that's part of the completions spec

though really I wish the dollar value was part of the completions spec..

[formal elk]

it seems from docs, its a auto cache, good to hear, so its best bet to choose only alibaba provider if using openrouter, assuming cache hit will work and handle.

[solar token]

It doesn't work, already tried it

Toven confirmed they dont have cache support rn for alibaba

[winter shore]

I am testing it now and can't get it to trigger

on their end

[torpid edge]

The unit price of cached_token is 40% of the unit price of input_token

not exactly great

[tacit tangle]

https://x.com/dzhulgakov/status/1948176962289360911

Dmytro Dzhulgakov (@dzhulgakov)

Qwen team keeps shipping: Qwen3 Coder 480B is live on @FireworksAI_HQ - on par with Sonnet 4 for coding! 🤯

https://t.co/jXADZh3oXp

Quick impressions:
• very strong agentic coding performance: SWEBench, Aider-Polyglot and other benchmarks are at the level of Claude Sonnet 4!

X · Jul 24, 2025, 12:22 AM

[dull sundial]

I obviously cross-test on unexpected results, same experience on alibaba

[dull sundial]

would love to know someones experience who uses coding agents on a daily basis, since #1 my bench isn't a coding specific bench (very little tasks in fact) and #2 doesn't use agentic workflows at all (also I don't plan to add it since I only want to test/evaluate what I am knowledgeable in).

[broken crane]

I will test it soon, but my evals are also not agentic. I could add some tool call evals into it. I think it will be helpful.

[astral ingot]

https://forgecode.dev/blog/kimi-k2-vs-qwen-3-coder-coding-comparison/
🤔

[broken crane]

Oh damn. Looks like qwen3 coder is no good.

[mystic dome]

do we have comparison of kimik2 vs claude4 sonnet?

[broken crane]

I have it here #1393208374769750227 message

[dull sundial]

I am changing stuff daily on my projects, kimi is good but sonnet is just another tier.. such a good model to work with

[brittle adder]

I drank the opus juice and I can’t go back

[dull sundial]

yea well opus is overkill for most stuff, but the absolute pinnacle imo

[brittle adder]

Valid.

[dull sundial]

opus did my entire chess replay system in 10 minutes work. its beautiful

3 prompts and 2 minor ui fixes. wouldnt have been possible 1 year ago without major headache

[brittle adder]

I’m just so glad to offload repetitive tasks anymore. Always check the code (even if it’s something it’s done hundreds of times), but holy crap have I saved so much dev time to focus on the work that I actually need/want to do it’s fantastic. Helluva time to be a software engineer!

[tight bane]

tl;dr YMMV

[broken crane]

Just finished testing Qwen3 Coder on my coding eval set, worse than Kimi K2. Will post full results soon.

[frozen relic]

"qwen3-coder-plus" and "qwen3-coder-480b-a35b-instruct" seem to have different prices on Alibaba. Are they actually the same model?

[tacit tangle]

Not sure but I saw this...

"qwen3-coder-plus supports context cache, which can reduce the cost of input tokens"

Also

https://www.alibabacloud.com/help/en/model-studio/qwen-coder

Says it has the same performance so im guessing same model

[broken crane]

Yeah my evals don't have tool call yet (it just tests that model supports tool call). I'm thinking of how to design an eval for tool call. Likely need to involve multi-round message and loops.

Some models can do multiple parallel calls in one response, which speeds up things.

I think for starter I'll test parallel tool call support and how many rounds it takes to get the result.

[broken crane]

A bit late, but finished testing Qwen3 Coder on my personal evals. Strong second in the open-source field. Qwen3 Coder outperforms DeepSeek V3 (New), but remains a step behind Kimi K2.

Observations:

• On standard medium-level tasks, Qwen3 Coder is among the best. It matches premium models in producing correct, concise code for markdown cleaning task.
• For more complex or formatting-sensitive challenges, like benchmark visualizations, it can lag behind due to rigid output formats or missing polish in its visual code.
• The main shortfall is in logical reasoning for uncommon programming patterns, such as advanced TypeScript narrowing, where it falls short alongside almost all open LLMs.
• Instruction-following is not particularly good, Qwen3 Coder tends to output more verbose blocks for "output only diff" tasks, similar to Kimi K2.

Overall better than DeepSeek V3 (New), but worse than Kimi K2 and Gemini 2.5 Pro. Not close to top model performance, such as Claude 4 models and GPT-4.1

Full evaluation blog post: https://eval.16x.engineer/blog/qwen3-coder-evaluation-results

[unkempt mist]

Looks like Cerebras is prepping a launch of qwen coder on their platform

[granite temple]

How does the pricing work here for the Alibaba providers?
Input $1.50 to $4.50
Output $7.50 to $22.50

What is the criteria for the price spike?

[rotund python]

Pricing matches here: https://www.alibabacloud.com/help/en/model-studio/qwen-coder

Toven know you're busy but there's no way to see the tiered pricing on mobile, and tap on android/brave does not bring up a tooltip

[neat jasper]

https://tenor.com/view/cleaning-chores-ironing-spongebob-squarepants-gif-4552991

Toven rn

[winter shore]

🫠

[granite temple]

thank you. There's also no way to check this through the API, like "max price" or something like that. At least a "warning" tag somewhere, meaning you should check the website before anything.

[inner patio]

qwen coder seems to have inherited the worse traits of both claude & gemini.
the claude yes man behaviour combined with the gemini2.5-06 "the user is wrong I will fight the user" behaviour. if qwen thinks something must be done it will do it no matter what. instructions and rules be damned. "you are asking this but I think you are asking me something else" No. im not. do what I asked. I understand this file is important to you, but theres a made up error I hallucinated so im going to remove it. I know you told me not to. great model overall but holy fuck is this kind of behaviour infuriating

[brisk glacier]

Seems like Cerebras are hosting this model now: https://inference-docs.cerebras.ai/models/qwen-3-480b

can't see it on OpenRouter: https://openrouter.ai/provider/cerebras

Why's that ?

[astral ingot]

they probably havent added it yet

[astral ingot]

@winter shore

[brisk glacier]

2k tok/s for $2/$2

Not too shabby

[astral ingot]

been loving kimi with groq, this might be a good alternative

[brisk glacier]

I wonder if it can be fixed with a better/adjusted system prompt 🤔

https://x.com/CerebrasSystems/status/1951339880325456159

Cerebras (@CerebrasSystems)

🟧🟨QWEN3 CODER is LIVE on Cerebras 🟨🟧
2,000 tokens/s - 20x faster than Sonnet
0.5s time-to-full-answer
131K context
$2 per M input/output tokens
Available in Cline, Windsurf & more

X · Aug 1, 2025, 5:50 PM

[solar token]

Is cerebras added for qwen 3 coder?

@winter shore can you add it? Seems like they added it on their site

[winter shore]

i am out today but team is aware

[solar token]

Bet tysm

[winter shore]

👀

[solar token]

Nice Ill keep an eye for when its up

tysm brother

@winter shore so just tested it, seems like it return error that toolcall failed, idk if something is miss configurated but yeah just letting you know.

[brisk glacier]

does the API take a moment to refresh?
I limited the provider to just cerebras and I'm getting this error: "No endpoints found for qwen/qwen3-coder."

[pulsar laurel]

I hold out hope they'll answer questions about quantisation, but I expect they won't - they've ignored everyone asking so far

OK I take that back. They ignored questions for the other releases, but for this one they say fp8. Nice https://xcancel.com/CerebrasSystems/status/1951358565563986279#m

[rotund python]

https://www.cerebras.ai/blog/introducing-cerebras-code they now have a max plan lol

[pulsar laurel]

With very misleading limits https://old.reddit.com/r/LocalLLaMA/comments/1mfeazc/cerebras_pro_coder_deceptive_limits/

[brisk glacier]

Yeah 7.5 M tokens / day is pretty disappointing

[formal elk]

good article, it almost makes me buy it, i know it , its sounds too good to be true.
plus its FP8

[dull sundial]

how so? Cerebras api is $2 mtok, so 7.8 million tok would be $15 (per day), on a $50 monthly plan. if we adjust for profit margin, maybe that's half to break even, still uses up the monthly fee in less than a week on a power user. sounds more like a hyper-inefficiency problem to me, you really don't need to feed 50k context on every edit.

[brisk glacier]

I mean, sure, you make a fair point, but if this is to be considered as a Claude Max competitor, this isn't it..

[dull sundial]

never used claude max, so wouldn't know. from a pure numbers standpoint, the cap seems quite generous, but I don't know the competing offers because I don't code like this.

[brisk glacier]

to be clear, I don't "vibe code" and send abnormal amounts of context for simple stuff. But given the "agentic tools" (e.g: opencode) we have today, those limits will run out fast.

[broken crane]

Claude Code is very generously subsidized now, even after the rate limit changes. I'm getting $10 value daily out of $20 monthly subscription.

[astral ingot]

theres no tool support on cerebras via openrtouer?

[fickle mantle]

best free model to use in opencode? kimi k2 small context but better than qwen 3 coder maybe. any others

[mossy pawn]

GLM4.5 is quite good

[rocky blaze]

how good

[fair breach]

They changed it to 24 million so apparently their profit margin is pretty healthy serving this…

[fickle mantle]

has it free api

[alpine flax]

Qwen3 Coder was the only open source model I tested that succeeded in implementing playable Monopoly.

I hope I don't get banned for self promo.
https://x.com/Avix_G/status/1952698267504193761

[uneven flint]

Qwen3 is not bad in opencode for me.

[slim vigil]

How come the free version has been removed today from OR? I had been using for a few days and was very happy

[frigid ginkgo]

Is it really gone? 😭

[slim vigil]

I found out when trying today in Cline. Went and checked the website and indeed it's not there anymore. :S Is there somewhere they inform about additions/removals?

[lethal yacht]

U expect a free provider to keep hosting a 480b model for ever?

its like 20 cent input and 80 cent output for 1M tokens

just pay

[slim vigil]

please spare me the morals

I'm just asking about the info

[lethal yacht]

be happy what u got, if its gone its not longer free pretty simple

[slim vigil]

thanks, so helpful

[plush badge]

atleast for additions you can check #new-models . I dont know for deletions.

[slim vigil]

do models get pulled from OR without notice then? Paid as well I guess.

[hoary prairie]

Free ones can go away at any time, there is no guarantee with them

[fickle mantle]

at the least there are free alternatives to do all the tasks with tools

[slim vigil]

free version is back!! :DDD

oh but only 32K context lol. Pretty unusable

[fresh onyx]

I think glm4.5 is better this time in my setting

[earnest dagger]

https://x.com/Alibaba_Qwen/status/1953835877555151134

Qwen (@Alibaba_Qwen)

💡 You get 2,000 free Qwen Code runs every day!

Run this one simple command:
npx @​qwen-code/qwen-code@latest
Hit Enter, and that’s it!
🚀 Now with Qwen OAuth support — super easy to use.
Try it now and supercharge your vibe code! 💻⚡
Github：https://t.co/8ITh20WTbV

X · Aug 8, 2025, 3:09 PM

[torpid edge]

But that's not for international users, no?

[earnest dagger]

Based on the replies, I'm guessing it's outside of China. Hopefully 🤞

[winter shore]

Quick note: Free Qwen3 Coder will be coming back today (with rate limits) thanks to Jon and Chutes. Say thanks in #1364073067713925161 !

[gritty whale]

Yayy!

[rocky blaze]

thanks jon and chutes!

[gritty whale]

https://chutes.ai/app/chute/e1026381-f55b-5a89-b74a-b579e073c420

There seems to be an fp8 version of this model that costs 1 cent in and 6 cents out

Toven, can you look into this?

This version of the model is wayyy cheaper than what's already on OpenRouter, and it supposedly is the exact same version as the model on OpenRouter right now

[winter shore]

that's like someone's private deployment

and it's not active / being used

idk how that works. but not gonna add it

look - 2 runs in 7 days, cold, and "Ronx123"

[gritty whale]

Mhm

[winter shore]

yeah idk what this is hahah

[gritty whale]

So only things made by Chutes are official/will be added on OpenRouter?

[winter shore]

the ones that are "Chutes" in the top there are the ones we'll add

I don't know what this other stuff is

[gritty whale]

👍

[winter shore]

It looks like other people can deploy models

but at our scale, we can't really add them

[calm kite]

Since last night, the model has been unstable, receiving rate limit messages. Looking at the uptime chart, I've noticed that I'm not the only one experiencing this issue.

[winter shore]

free models have no uptime guarantees, if there’s too much demand yes the uptime will suffer

it’s doing >10B free tokens per day

that’s ~$3,000+ dollars a day for free

[calm kite]

I see, but it was doing well before. like almost 100% uptime. that is why i was wondering if something was wrong with it.

[earnest dagger]

https://x.com/petergostev/status/1956360226845716842

Peter Gostev (@petergostev)

Anthropic and Google are losing coding share in recent weeks, according to @OpenRouterAI data. This isn’t really because of GPT-5 (it has 20% in this dataset.

Caveat: this reflects only a small slice of the market. It

X · Aug 15, 2025, 2:20 PM

[hollow delta]

is qwen3 coder that good or is it the free variant just making it really attractive for the price of free?

[earnest dagger]

Good question. Idk how to see how many tokens are processed by each providor though :/

https://openrouter.ai/qwen/qwen3-coder/activity
anyone?

[earnest dagger]

I checked free vs paid, and its 14 vs 9 billion prompt tokens today

So the paid is still at higher usage, just given today.

[dusty tiger]

Have you seen this before? What can I do about it? I'm paying for the service, so I've chosen the paid option.

[gritty whale]

that means too many users are using Qwen3 with that specific provider. That provider is rate limiting OpenRouter

you are not charged for the requests that are rejected

use a different provider

[dusty tiger]

Thank you for the quick response. I'm using the provider Baseten.

[gritty whale]

I've never seen Baseten rate limit before, but if your request is being rejected while using Baseten, using a different provider will fix it

[dusty tiger]

ok, thanks

[hollow delta]

you could also provide a list of providers as fallback if you're using the api

afaik

[dusty tiger]

I have strictly regulated which providers I use, as I've had disappointing experiences with some providers—for example, insufficient quantization or excessively high costs. That's why I only maintain a positive list of providers and have now expanded it with an additional provider. I hope this will now work.

[astral ingot]

i think someone already mentioned this but i got an empty response from gmicloud

[hard bramble]

What providers do you prefer? I'm trying to assemble my own list of whitelisted and/or blacklisted providers

[pliant quiver]

@dusty tiger I'm interested as well, which one do you use and do you use it with tools?

[dusty tiger]

@hard bramble @pliant quiver

[hard bramble]

Nice, thanks! You have a good experience with Chutes? I've found it to be hit-or-miss.

[dusty tiger]

I didn't have the best experience with Qwen3 Coder yesterday in general. I kept hitting the limit. I then switched back to a commercial model.

[thick jacinth]

when will its api come out?

[hoary kayak]

When will it appear on openrouter?

[gritty whale]

It already has

[hoary kayak]

oops, sorry(

[thick jacinth]

can you give me the exact name or link i am having trouble finding it

[gritty whale]

https://openrouter.ai/qwen/qwen3-coder

[thick jacinth]

thanks I mixed it up

[half agate]

Hi!
Getting some strange results in Chat using specific providers:
Deepinfra - No endpoints found for qwen/qwen3-coder.
Together - No endpoints found for qwen/qwen3-coder.
Baseten - error
Parasail - working
Fireworks - working
NovitaAI - No endpoints found for qwen/qwen3-coder
AtlasCloud - working
Phala - working
etc

Am I doing something wrong or some providers changed endpoint names?

[modest stream]

same issue here, just I have a funded today my openrouter account.
I would like to request access to the "qwen/qwen3-coder" model, but it doesn’t show up in my available models, if there are current regional restrictions?

[dusty tiger]

I tried qwen3 yesterday for a UI bug. First of all, it took several attempts to correct the error. And secondly, it was significantly more expensive than even GPT 5 high. Oh, and it also got stuck in an endless loop at one point.

[broken crane]

is alibaba/opensource the same as old alibaba/plus?

[broken crane]

https://x.com/GosuCoder/status/1958903833179947087

GosuCoder (@GosuCoder)

With over 200 million tokens over the last few days put into Qwen 3 Coder provider testing. I finally have results I can share.

1. I think we need to be able to multi-select providers in any ai coding tool.

2. API reliability was a lot bigger of a problem with Qwen 3 Coder than

X · Aug 22, 2025, 2:47 PM

[ivory tangle]

Are there no providers with prompt caching?

[winter shore]

no, plus is coder-plus in alibaba api

[broken crane]

Oh. Why was it removed?

[winter shore]

it's not the same model as the open source weights

it will get it's own model page soon

it's the proprietary closed source weights of the qwen3 coder weights

[broken crane]

Got it

[drifting aurora]

@dusty tiger idk if you are still using it, but I really like using Cerebras for qwen 3

[dusty tiger]

Thanks for your message. I tried again yesterday and received inline code in Kilo Code. So the code was not written to a file but to the Kilo Code chat box.

[drifting aurora]

I am not sure how it would work with kilo code and other ai coding softwares, but I mostly use it with api. Also I tried to give Kimi k2 access to tool calling and make it call qwen 3(Kimi for the main one, because it's better for tool calling), but I couldnt get it to edit files, it only made new ones

But for you use, it's prolly worth to wait for their code pro/max plans to restock, they work with cline

[dusty tiger]

yeah, I will wait

[glacial dragon]

Алло нахуй

[dusty tiger]

English please

[glacial dragon]

Where can I use the LLaMA2 uncensored model?

[drifting aurora]

Not on here

[glacial dragon]

Нихуя не понял

[drifting aurora]

Stop fucking speaking russian

[glacial dragon]

🇷🇺

[drifting aurora]

fucking retard

[glacial dragon]

💩

[gritty whale]

There's no need for that language here

[valid bluff]

zzz

[ivory tangle]

Anyone know a provider that supports prompt caching?

[thick jacinth]

nope ., open source models rarely have prompt catching

you can try deepseek v3.1 from deepseek directly

[ivory tangle]

Thanks, I use deepseek V3.1, GLM4.5 and Kimi K2, and was wondering why my Qwen bill was always the highest

Guess it’s the only one not prompt caching

[thick jacinth]

yes , i wasted like 2$ before realizing that , you can save upto 90% depending on what you chat about with prompt caching

[ivory tangle]

Yeah, might swap to requesty for all of my qwen3 coder tasks since they have prompt caching via alibaba, then keep using OpenRouter for Kimi/GlM/deepseek

[winter shore]

we'll add coder-plus

which is the cache enabled model

it's not the same model as the other qwen3 coder though

[winter shore]

random Q, did you get a DM telling you requesty had caching?

[ivory tangle]

No I did not, there is a hackathon roo x requesty, and they give free credits if you participate, so when I went there I saw Qwen had caching. I’m assuming you are checking to make sure no-one is trying to poach or users via DM

And yeah I know they are dif models, I think even tho it’s more expensive on a per M token bases, i think it will be cheaper bc of caching, and I will cap the context at 128k or something, once you guys have it added I will use it here, much prefer having everything via one balance

[glad reef]

Is it added?

/ How would I access it?

[ivory tangle]

I wonder if its taking a while because they have price tiers instead of a simple input and output price

[unique musk]

hi

[hoary raven]

By the way guys, if anything happened with Qwen again, try asking in the Alibaba Cloud Developer Community Discord. Lara’s great at bringing the Qwen team in for quick answers.

[gleaming fulcrum]

@winter shore just wondering about Alibaba provider not being in the "exacto" endpoints? Surely the Alibaba API is a reference quality implementation?

fwiw I get better and more consistent results from qwen/qwen3-coder-plus (which only has Alibaba as a provider for some reason), than other qwen3 coder endpoints.

[gleaming fulcrum]

in general though exacto is a massive step in the right direction

[viscid flint]

@gleaming fulcrum the plus model is a different proprietary model and likely uses different weights from the open source ones. That’s probably why it’s better.

[winter shore]

their open source endpoint on qwen3coder is not really that great

in our real world data and in benchmarks

the decisions for exacto are based on data, not on wheter it would make sense to have it

the other models have the model author endpoints in exacto because they're good endpoints

[gleaming fulcrum]

fair enough. I didn't realize that qwen3-coder-plus and open source endpoints were different.

[rich field]

rip

[gritty whale]

Why is Cerebras taking these models?

Are they strapped for GPUs or something?

[rich field]

im not sure. they could keep both but for a reason they need to remove it

really weird

it was nice with the speed

[gritty whale]

Mmm

[neat jasper]

They don’t even run GPUs don’t they