Qwen3-coder release

[nav1dn]

Hi, It would be great if Qwen3coder get released for coding, just like qwen2.5-coder. it saves our time.
Just make sure it trained with fresh/updated high quality codes, especially for system-programming-languages.

[jklj077]

it is currently in the plans.

[wienans]

That’s nice hopefully it supports good FIM too.

Lately there is a drop in FIM publicity :/

[pranayagra]

Is there any rough estimate on when it will be released? Or is it far out (4+ months)?

[g0t4]

@wienans yup Qwen2.5-Coder is a rockstar with FIM... so damn fast and good... plus Qwen3's new Embeddings and Re-ranking models :)...

[leekahwai]

I hope Qwen3-coder can touch more in c and c++ space. Current AI models are all tuned for web and desktop development. But a lot of technology we use today are still embedded in smart devices and most companies can't switch to rust due to dependencies on legacy hardware components.

[wienans]

Yeah embedded systems will not switch to rust or anything else any time soon. The switch to c++ isn’t even really done most things are still build in plain C. And I think that will never change because embedded product lifecycle is so long in comparison.

[g0t4]

fine tune Qwen2.5-Coder on some c/c++ code? perhaps even your own code bases?

[pinturic]

Is there any eta for qwen3-coder?

[jackytank]

hi is there any update regarding the release of Qwen3-Coder tks.

[pai4451]

Interested in the release of Qwen-3-Coder model

[Luc45]

[Teeeto]

Will be released right after half-life 3 :) Seriously, its highly anticipated.

[antonpetrovmain]

What is the planned max context size for the new model :) ?

[spgunarath]

This will be the most widely used open model ever. Alibaba Cloud should provide ultra-fast endpoints like Groq—this would drive even greater adoption and help them secure more funding.

[matthieuHenocque]

It's been a long time since Qwen released a Coder model... (I'm trying something)

Anyway, thanks guys for your amazing work. Can't wait to put my hands on Qwen3-coder!

來自法國的問候！

[n-WN]

Live

It's been a long time since Qwen released a Coder model... (I'm trying something)

Anyway, thanks guys for your amazing work. Can't wait to put my hands on Qwen3-coder!

來自法國的問候！

[fernandaspets]

Please upload Qwen3-Coder to Hugging-face

[ernestp]

It's on Hugging Face, enjoy )

https://huggingface.co/Qwen/Qwen3-Coder-480B-A35B-Instruct

[wienans]

Will the smaller variants support FIM, maybe a dedicated Fim model?
All people go crazy on agentic coding but supporting the programmer while writing code is I think similarly important, as we can not hope to do all code via agentic coding especially as it is sometimes overkill.

[pai4451]

Also interested if a smaller size Qwen3 Coder will be available too?

[g0t4]

Qwen3-Coder-30B-A3B is working well in my testing, on par w/ to Qwen2.5-Coder-7B though I haven't done benchmarks, just in my daily usage

[jackytank]

Jesus christ Qwen3-coder 480B so good guys, It would be great to have smaller FIM qwen3-coder (1.5b or 3b is nice for FIM) to self-host with Ollama/LmStudio in my potato laptop

[g0t4]

try it out! Qwen3-Coder-30B-A3B

[ernestp]

Smaller model is already here - https://huggingface.co/Qwen/Qwen3-Coder-30B-A3B-Instruct

[wienans]

But can it do fim? At least it is not in the model card?

[pai4451]

@ernestp Does it support fill in the middle? I want to find a code completion model

[reneleonhardt]

Wow, thank you very much, the smaller model is super fast 🚀

Would it be possible to minimize the coding benchmark gap (i.e. 51.6 to 69.6) to 480B-A35B significantly without getting multiple times bigger or is there no reasonable "middle ground"? 😄

[wienans]

https://github.com/QwenLM/Qwen3-Coder?tab=readme-ov-file#fill-in-the-middle-with-qwen3-coder

It looks like it does support FIM @pai4451
Didn’t try yet

[reneleonhardt]

But can it do fim? At least it is not in the model card?

Config contains FIM:
https://huggingface.co/Qwen/Qwen3-Coder-30B-A3B-Instruct/blob/main/tokenizer_config.json#L133

[pinturic]

Any plan to have smaller models? I am using at the moment in my machine this https://huggingface.co/Qwen/Qwen2.5-Coder-7B-Instruct and it is doing really great! it would be really great to have the qwen3-coder version

[g0t4]

FYI speed wise Qwen3-Coder-30B is running as fast as 2.5-Coder-7B (which I use extensively for FIM)... Memory demands are bigger obviously! What hardware are you using?

Spec decoding would be another good reason to have a smaller Qwen3-Coder. I use Qwen2.5-Coder-7B judge with 2.5-Coder-0.5B draft model... and wow that combo is fast... 300 tokens/sec no problem. I tried using 2.5-Coder-0.5B as the draft model for Qwen3-Coder but got bad results (might need to adjust some parameters)...

Separately, getting a BASE model (non-instruct) for Qwen3-Coder would also be useful given the Instruct models tend to ramble on FIM requests. Base models are more likely to stop as one would expect.

What type of requests are you sending?

[pinturic]

My use case is this:

1. I have a mac os with 32 Gb of ram
2. I am using lm studio with Qwen2.5-Coder-7B-Instruct to generate unit tests code
3. I am using lm studio with Qwen2.5-Coder-7B-Instruct with cursor as an agent

[ernestp]

I'm using hf.co/unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF:Q2_K_L in Ollama on 32GB Macbook without any issue

I'm also using

export OLLAMA_KV_CACHE_TYPE=q8_0
export OLLAMA_FLASH_ATTENTION=1

[g0t4]

@pinturic have you tried Qwen3-Coder for FIM with one of these:

Qwen3-Coder-30B-A3B-Instruct-MLX-5bit (or 4/6 bit? they're range from 17 to 24GB in size)

I've been using Qwen3-Coder for both FIM and Chat and it's working well

[pinturic]

It is nice! but I was hoping to have something smaller

[reneleonhardt]

A Qwen3-Next-coder version of the new super-fast and compact Qwen3-Next model would be awesome! 🚀

https://huggingface.co/Qwen/Qwen3-Next-80B-A3B-Instruct

• Qwen3-Next-80B-A3B-Base outperforms Qwen3-32B-Base on downstream tasks with 10% of the total training cost and with 10 times inference throughput for context over 32K tokens.
• Qwen3-Next-80B-A3B-Instruct performs on par with our flagship Qwen3-235B-A22B-Instruct-2507 on certain benchmarks, while demonstrating significant advantages in handling ultra-long-context tasks up to 256K tokens.

https://qwen.ai/blog?id=e34c4305036ce60d55a0791b170337c2b70ae51d&from=home.latest-research-list

• In inference, especially with context lengths over 32K tokens, it delivers more than 10x higher throughput
• The Qwen3-Next-80B-A3B-Instruct performs comparably to our flagship model Qwen3-235B-A22B-Instruct-2507, and shows clear advantages in tasks requiring ultra-long context (up to 256K tokens). The Qwen3-Next-80B-A3B-Thinking excels at complex reasoning tasks — outperforming higher-cost models like Qwen3-30B-A3B-Thinking-2507 and Qwen3-32B-Thinking, outperforming the closed-source Gemini-2.5-Flash-Thinking on multiple benchmarks, and approaching the performance of our top-tier model Qwen3-235B-A22B-Thinking-2507.

https://qwen3next.org/

Is commercial use allowed?

Yes. Qwen3-Next is open-sourced under Apache 2.0. You can self-host on-prem/VPC or use cloud services, modify and redistribute derivatives with license notices, and integrate into products with your governance in place.