Qwen/Qwen3.8-2.4T-A95B

Qwen has released the post-trained model weights and configuration files for Qwen3.8-2.4T-A95B in Hugging Face Transformers format. The model is compatible with major inference frameworks such as vLLM, SGLang, and TokenSpeed.

Qwen3.8 is the latest generation of the Qwen open model family, with performance improvements focused on coding, specialized tasks, research, and long-term agent work. It has enhanced autonomous planning that reflects environmental feedback and task completion reliability, and allows users to adjust reasoning depth via reasoning_effort and preserve the reasoning context of previous messages via preserve_thinking.

• Total parameters: 2.4T
• Active parameters: 95B
• Architecture: 10 routing experts and 1 shared expert activated out of 512 MoE experts
• Base context length: 262,144 tokens
• Extendable context: up to 1,010,000 tokens
• MTP (Multi-Token Prediction) training support

In benchmarks, Qwen3.8-Max recorded PaperBench 93.0, IFBench 82.8, HealthBench 60.2, and MRCR v2 256K 92.9. In coding agent evaluations, it recorded Terminal Bench 2.1 86.6, SWE-bench Pro 67.7, and QwenSWEBench 80.7.

For users seeking managed inference, the Qwen Cloud API is also provided. The official commercial version, Qwen3.8-Max, additionally supports vision input, non-thinking mode, a default 1M context, and built-in tools.