Qwen3.8 ships a 2.4-trillion-parameter open-weight model that activates 95B

The Qwen team released Qwen3.8, a mixture-of-experts model with 2.4 trillion total parameters but only 95 billion active per token. That sparsity is the point: you get the capacity of a very large model while paying compute closer to a 95B one. The architecture is hybrid, with 92 layers that alternate Gated DeltaNet and gated attention blocks feeding into the MoE, and 512 experts with 11 active per token (10 routed plus one shared).

Context runs to 262,144 tokens natively and stretches to about 1 million with extension. The model card reports 80.7% on QwenSWEBench and 92.6% on GPQA Diamond, which puts it in range of Claude Opus 4.8 and GPT-5.6 Sol on those tasks. Two training changes are new here: multi-token prediction, and a reasoning_effort control that lets you set thinking to low, medium, or xhigh. The model always thinks before answering rather than offering a non-reasoning mode.

The weights are on Hugging Face under a "qwen3.8-max" license, so read the terms before commercial use rather than assuming the permissive license earlier Qwen models shipped with.

Why it matters

An open-weight model at this capacity gives teams a frontier-class option they can host themselves, but the sparse design and custom license change the calculation: budget for the memory to hold 2.4T parameters even though only 95B compute per token, and check the license terms before you build on it.