2.4T total / 95B active (MoE)
262,144 tokens native; extensible to 1,010,000 tokens
Qwen3.8-Max License
Qwen's first open-weight Max-class release and the largest model in the Qwen open family at launch, provided as a post-trained causal language model with 2.4 trillion total parameters and 95 billion activated per token. Built on the Qwen3.5 architectural foundation, the model has 92 layers arranged as 23 repetitions of three Gated DeltaNet-plus-MoE blocks followed by one gated-attention-plus-MoE block. Its sparse architecture contains 512 experts with ten routed experts and one shared expert active per token, a hidden dimension of 8,192, a padded 248,320-token vocabulary, grouped-query attention with 64 query and four key-value heads, and multi-step Multi-Token Prediction training. Qwen positions the release for coding, professional work, research, and long-horizon agentic execution, emphasizing stronger autonomous planning, environment-feedback handling, harness compatibility, and reliable end-to-end completion of complex tasks. The checkpoint supports adjustable reasoning depth through xhigh, medium, and low reasoning_effort levels and retains historical reasoning using preserve_thinking. It is text-only and requires thinking mode: every response begins with reasoning enclosed in think tags, and non-thinking generation cannot be enabled. This distinguishes the downloadable checkpoint from the hosted Qwen3.8-Max service, which is based on it but adds vision input, non-thinking support, built-in tools, and a default one-million-token context configuration. Published Qwen evaluations report 86.6 on Terminal Bench 2.1, 67.7 on SWE-bench Pro, 73.5 on FrontierSWE, 93.0 on PaperBench, 81.9 on WideSearch, 92.6 on GPQA Diamond, 82.8 on IFBench, and 92.9 on MRCR v2 at 256K. The approximately 4.89 TB release is distributed in Hugging Face Transformers format with deployment support documented for SGLang, vLLM, and TokenSpeed.