Alibaba ships Qwen3.8-Max open weights, but with a revenue-share license
Companies making more than $50 million a year must obtain a separate paid license to use the model commercially, breaking from the free Apache terms Alibaba used before.
Original cover art, generated for this story. THE VISSION does not republish third-party press imagery.
- Alibaba released open weights for Qwen3.8-2.4T-A95B, a 2.4-trillion-parameter mixture-of-experts model with 95 billion active parameters, according to the model's own Hugging Face listing.
- The published license requires a separate paid license once a licensee's aggregate revenue exceeds $50 million in any 12-month period while running the model as a service.
- The open-weight release is text-only and requires "thinking mode" for every interaction, missing the vision input and full 1-million-token context available through Alibaba's hosted API version.
Alibaba released open weights for Qwen3.8-2.4T-A95B — a 2.4-trillion-parameter mixture-of-experts model with 95 billion parameters active per token — but attached a license that breaks from the freely licensed Apache 2.0 terms the company used for earlier Qwen releases, according to the model's Hugging Face listing and its published license file.
The license requires anyone running the model as a paid service, or as an "AI Work Assistant" business, to obtain a separate commercial license once their aggregate revenue, or their affiliates', exceeds $50 million in any consecutive 12-month period. Products with more than 100 million monthly active users or $20 million in monthly revenue must also prominently display the model's name. The open-weight release is text-only and requires "thinking mode" for every interaction, and its native context window is 262,144 tokens — short of the full 1-million-token context and vision input available through Alibaba's hosted Qwen3.8-Max API.
The move follows a similar revenue-share licensing shift Moonshot AI made with its Kimi K3 model. Alibaba's own benchmark figures for the open release include a Terminal-Bench 2.1 score of 86.6 and a SWE-bench Pro score of 67.7 — figures from the company's own model card that have not been independently reproduced.
"Open weights" increasingly does not mean free to build a business on — a well-funded startup can no longer assume it can scale a product on a downloaded model without eventually owing its creator a licensing fee, which changes the actual cost calculus of choosing open weights over a metered API in the first place.