Alibaba open-sources Qwen 3 Coder 480B after topping SWE-bench leaderboard

Read the original article →

What happened

Alibaba's Qwen team released Qwen 3 Coder 480B as open weights on July 27. The model topped the SWE-bench Verified leaderboard at 76.8 percent before release.

The MoE architecture activates 35B parameters per token and supports 256K context with native tool calling. Weights are on Hugging Face under Apache 2.0. Alibaba also offers hosted inference through DashScope and Model Studio.

Why it matters

An open-weight model leading SWE-bench challenges the assumption that closed frontier models always win on coding. The MoE design also makes inference cheaper than a dense 480B model would require.

Apache 2.0 licensing removes friction for commercial deployment, which matters for teams that rejected Llama's license terms.

MintedBrain take

Leaderboard leadership is a snapshot, not a guarantee on your codebase. Self-hosting a 480B MoE still needs serious GPU capacity, so try the hosted API first before planning on-prem deployment.

References

Discussion

  • Loading…

← Back to News