OPEN-SOURCE
Apr 22, 2026Alibaba releases Qwen3-Max-Reasoning under Apache-2.0
Qwen3-Max-Reasoning posts top-three scores on AIME and SWE-Bench Verified. Weights are open, license is Apache-2.0, and the chat template ships with the inference repo.
Alibaba’s Qwen team published Qwen3-Max-Reasoning, the reasoning-mode variant of the 480B-parameter Qwen3-Max model released earlier this year. Reported scores: 92.4 on AIME-2025, 71.8 on SWE-Bench Verified, 88.1 on GPQA-Diamond. That places it inside the top three on every major reasoning benchmark, alongside Claude Opus 4.7 and Gemini 3.1 Pro Thinking.
The notable detail is licensing. Where most Western reasoning models remain closed-weight, Qwen3-Max-Reasoning is Apache-2.0, weights and chat template both. Inference at full precision needs roughly 8 H100s; the team also published a 4-bit AWQ quantization that runs on a single A100 80GB at modest tokens-per-second.
For teams wary of provider lock-in, this is the cleanest open alternative to closed reasoning models so far. The Qwen team also confirmed Qwen3.5 is in training, with a stated target of “significantly improved long-context reasoning”.
Source
github.com/QwenLM