Alibaba open-weights Qwen3.8, its first Max-class model released with open weights

Alibaba’s Qwen team published open weights for Qwen3.8-2.4T-A95B on August 12, 2026, the first time it has released a model in its top Max class with open weights rather than keeping it behind an API.

That is the change: when Qwen3.8-Max first launched on August 3, 2026, it was a closed model. This drop makes the 2.4-trillion-parameter checkpoint downloadable, the weights now posted on Hugging Face.

The model is a mixture-of-experts system that activates 95 billion of its 2.4 trillion parameters per query, using a hybrid architecture that interleaves Gated DeltaNet and Gated Attention blocks feeding the expert layers. It carries 512 total experts with 11 activated per forward pass, 92 layers, and a native context length of 262,144 tokens extensible to about 1.01 million, Qwen said. A smaller companion checkpoint, Qwen3.8-27B, was released alongside it.

The move continues China’s open-weight push against closed U.S. frontier labs, giving researchers and enterprises a Max-class model they can host and fine-tune themselves.

The license is the fine print. The model card lists a custom “qwen3.8-max” license rather than a standard permissive one, so commercial terms and usage restrictions matter as much as the weights being public — open weights are not the same as open source. Qwen has not published head-to-head benchmarks against leading closed models with this release.