Alibaba’s Qwen team published the weights for Qwen3.8-2.4T-A95B on 2026-08-12, according to the news log in the official Qwen3.8 repository, with the smaller Qwen3.8-27B following on 2026-08-14. Both landed on Hugging Face and ModelScope. The team states plainly what is new here: for the first time, Qwen3.8 brings a Qwen-Max-class model to open release. Previous Max-tier Qwen models were API-only.
The flagship checkpoint is a mixture-of-experts model with 2.4 trillion total parameters and 95 billion activated per token, built on 92 layers that mix Gated DeltaNet with attention, a hidden dimension of 8,192, and 512 experts. Native context is 262,144 tokens, extensible to roughly 1,010,000. It is a text-generation causal language model that requires thinking mode for every interaction, with reasoning depth set through a reasoning_effort parameter offering xhigh, medium and low; xhigh is the default. Serving is supported on vLLM, SGLang and TokenSpeed. The license is not permissive: the model card lists a bespoke qwen3.8-max license rather than Apache 2.0.
Qwen3.8-27B is the companion release and is licensed Apache 2.0. It is a dense 27 billion parameter model with a vision encoder, native image and video understanding, the same 262,144 native context extensible to 1 million, and the same flexible thinking controls. Reported results include 73.0 percent on Terminal Bench, 61.7 percent on SWE-bench Pro, 84.3 percent on OSWorld-Verified and 81.9 percent on AndroidWorld.
Two things follow for anyone evaluating open models. First, the frontier-scale checkpoint is downloadable but not freely licensed, so the legal review matters as much as the benchmark review; the genuinely permissive option in this drop is the 27B. Second, a 2.4 trillion parameter model with 95 billion active parameters is not a self-hosting proposition for most organizations, so its practical effect is on the hosting market rather than on private deployment: third parties can now serve a Max-class Qwen, which puts price pressure on Alibaba’s own API.