Alibaba open-weights its largest model, a 2.4-trillion-parameter flagship
The mixture-of-experts model activates 95 billion of its 2.4 trillion parameters per token; the open weights are text-only, unlike the hosted version's vision input and larger context.
- Open weights & ecosystem
- Models & capabilities
- Notable
Alibaba published open weights for Qwen3.8-2.4T-A95B, ten days after it unveiled the model as a hosted API product on 3 August. At 2.4 trillion total parameters with 95 billion active per forward pass, this is the first time Alibaba has open-sourced a Qwen-Max-tier model — previously it kept its largest models closed while releasing smaller ones openly.
The model is a fine-grained mixture-of-experts design with 512 total experts, 11 activated per token (ten routed plus one shared), across 92 layers alternating gated DeltaNet and gated-attention layers. It natively supports 262,144 tokens of context, extensible to roughly 1 million, and Alibaba made it available on Hugging Face compatible with vLLM, SGLang and other standard inference stacks.
The open release is not equivalent to the hosted Qwen3.8-Max: the published weights are text-only and operate exclusively in “thinking” mode, while the API version additionally offers vision input, a non-thinking mode, a larger default context window and built-in tools. And at 2.4 trillion parameters, running the full checkpoint remains, in practice, a multi-node datacentre undertaking rather than something available on ordinary hardware — Alibaba had positioned its smaller Qwen3.8-27B, released alongside the August 3 announcement, as the realistic option for local deployment. The release nonetheless continued a pattern through 2026 of Chinese labs pairing their largest, most capable models with open licensing that most Western labs had by then abandoned.