Alibaba's Qwen3.8-Max open weights finally land

Alibaba published open weights for Qwen3.8-Max on Hugging Face on August 12, three weeks after previewing the model on its Token Plan and Qoder API in July. The checkpoint, Qwen3.8-2.4T-A95B, carries the full 2.4 trillion total parameters with 95 billion active, plus an FP8 variant, marking the first time Alibaba has open-weighted a Max-class model. It’s the first confirmation of a model Alibaba’s own team said in July would go open-weight “soon.”

The release comes with real trade-offs against the API version: it’s text-only rather than multimodal, it drops the 1-million-token context window, and it ships under a new custom Qwen3.8-Max License rather than a standard permissive license, with revenue-sharing terms Alibaba hasn’t fully detailed publicly. The smaller Qwen3.8-27B companion model promised alongside it has not shipped yet.

What it means for you

The open-weight version is meaningfully smaller in capability than the API model Alibaba benchmarked publicly, no multimodal input and a fraction of the context window, so don’t assume you’re getting the same model you’d get through the API just because the weights are downloadable. Before you self-host this for a production workload, get legal eyes on the revenue-sharing license terms specifically, that structure is unusual enough among open-weight releases to warrant its own review before you self-host it in production. It follows the same wait-and-see caution flagged when the preview first launched: treat vendor benchmark claims as marketing until the actual shipped artifact is in your hands, the same read worth applying to DeepSeek V4 Pro’s jump out of preview this same week.