Qwen3-Coder is Alibaba's dedicated agentic-coding model: 480B total parameters with 35B active, 256K native context extrapolating to 1M, and 69.6% on SWE-bench Verified — for much of 2025–26 the strongest pure-coding score in open weights under Apache 2.0.
It ships with Qwen Code (a CLI adapted from Gemini CLI) and slots into Claude Code-style agent workflows. The smaller 30B-A3B variant brings the same recipe to single-GPU setups, making the Coder line the most complete open-weight coding stack from one vendor.
Qwen3-Coder specifications
| Vendor | Alibaba (Qwen team) (China) |
|---|---|
| Architecture | Mixture-of-Experts, code-specialized |
| Total parameters | 480B |
| Active parameters | 35B |
| Context window | 256K tokens (1M with extrapolation) |
| Modalities | text |
| License | Apache 2.0 — permissive, patent-granting, free for commercial use |
| Release date | 2025-07-22 |
| Weights | Hugging Face · GitHub |
Benchmarks and reported results
| Benchmark | Result | Note |
|---|---|---|
| SWE-bench Verified | 69.6% |
Running Qwen3-Coder locally
480B total / 35B active: about 270GB at 4-bit. Widely hosted; locally it wants a multi-GPU node, though smaller Qwen3-Coder variants (30B-A3B) run on a single 24GB GPU.
What Qwen3-Coder is best for
- Agentic coding under the cleanest enterprise license (Apache 2.0)
- Repository-scale refactoring (256K–1M context)
- Single-GPU coding via the 30B-A3B variant
Frequently asked questions
Is Qwen3-Coder better than general models at coding?
On pure software-engineering benchmarks it led the Apache-2.0 field with 69.6% SWE-bench Verified. Newer generalists (Kimi K2.5, GLM-5.2, DeepSeek V4) have since matched or passed it, but it remains a focused, widely-hosted coding specialist with an unbeatable license.