Alibaba puts a 2.4-trillion-parameter model on Hugging Face
Qwen3.8-2.4T-A95B — a 2.4-trillion-parameter sparse mixture-of-experts model with roughly 95B active — shipped as open weights on 12–13 August. It is the first Max-tier Qwen ever made downloadable, and it is a text-only checkpoint without the vision and 1M-context capabilities of the hosted version.
Max-tier was the line Alibaba did not cross. The Qwen family has shipped open weights at every size for two years while keeping the flagship behind an API, exactly as every other lab does. That line is now crossed, and the parameter count is not a rounding error: 2.4 trillion total, roughly 95 billion active.
What shipped is narrower than what is hosted. This is a text-only checkpoint; the vision capability and the 1M-token context of the served model are absent. That is a meaningful asterisk on any "the frontier is now open" reading, and it is the kind of detail that gets lost between the announcement and the commentary.
The practical constraint is that almost nobody can run it. A 2.4T sparse model is a datacentre artefact, not something that fits the single-consumer-GPU story that has driven open-weight adoption all year. Its value is to organisations with serving infrastructure, and to researchers who can now inspect a frontier-scale model's internals directly.
Which makes the release partly a statement about who sets terms. The licence attached to it is the more consequential change, and the sibling model promised for the same week has not appeared at all.
explainX.ai — Qwen3.8-Max Open Weights Are Live (August 2026) → · LLM Stats — AI Updates Today (August 2026) — Latest AI Model Releases →