Qwen3-Coder review: Alibaba's open 480B model for agentic coding
FastMetal no longer offers Qwen3-Coder. This review is kept as a record of the model at its launch; what we serve today is on the models page, and the model's own page is Qwen3-Coder.
The verdict: Qwen3-Coder (Qwen3-Coder-480B-A35B-Instruct-FP8) was the open model to pick if you wanted to build coding agents and automation on weights you control. It is at its best on multi-step work that generates code, edits it and calls tools, and it trails the top closed models by very little on the standard agentic coding benchmark. For general chat and writing, a general-purpose model is easier to work with.
What it is
Alibaba's Qwen team released Qwen3-Coder in July 2025.
- Mixture of experts: 480B parameters in total, about 35B active per token
- Context window: 256K tokens natively, and up to 1M with extrapolation, according to Qwen
- License: Apache 2.0, so commercial use is straightforward
- Built for repository-scale work: the long context is there so the model can read a whole codebase at once
Alibaba also open-sourced Qwen Code, a command-line tool for handing development tasks to the model in plain language.
Agentic coding is the point
Qwen3-Coder is tuned for agent loops: using tools, running tests, and editing code over many steps, more than for line completion. The Qwen team describes it as leading open models on agentic coding, browser use and tool use, at a level comparable to Claude Sonnet 4.
The SWE-bench Verified figures OpenRouter shared at launch (500-turn setting) back that up:
| Model | SWE-bench Verified |
|---|---|
| Claude Sonnet 4 | 70.4% |
| Qwen3-Coder | 69.6% |
| OpenAI o3 | 69.1% |
| Kimi-K2 | 65.4% |
| GPT-4.1 | 54.6% |
Early developer reports matched the benchmark's picture. It does well at the iterative loop of opening the code, changing something and rerunning the tests, and the long context means it rarely loses track of a variable defined in another file.
Where it slips
Read the 69.6% with its setup attached. It comes from one harness (500 turns, OpenHands-style scaffolding); a standard configuration was reported at 67.0%. SWE-bench scores move with the scaffold, the turn limit and the tools available, so test it on your own workflow before trusting a number.
Hands-on reviews also found small logic bugs in otherwise good code. A short follow-up instruction usually fixed them, and that is the practical way to run it: expect a correction pass.
Its design leans toward development. For open-ended chat, prose and summaries, a general model such as Claude is easier to get good results from. It reads instructions and requirements written in other languages well enough, and it can comment code in them, but polished non-code writing is not what it was built for.
Good fit, poor fit
Good fit:
- Code generation, completion and refactoring
- Agentic automation that runs tests between steps
- Reading and changing large codebases
- High-volume coding work where token cost matters
Poor fit:
- Products built mostly on chat, writing or summarization
- Work where the quality of natural-language prose comes first
FAQ
Is Qwen3-Coder better than Claude?
At coding it comes close. On SWE-bench Verified (OpenRouter's figures, 500 turns) Qwen3-Coder scored 69.6% and Claude Sonnet 4 scored 70.4%. The scores move with the setup, so the practical answer is to match each model to the job. Our later head-to-head, Qwen3-Coder vs Claude Sonnet 5, covers how the two split the work.
How long is the context window?
256K tokens natively and up to 1M with extrapolation, according to Qwen. It is sized for feeding in a large codebase in one go.
Can I use it commercially?
Yes. Qwen publishes it under Apache 2.0.