Alibaba open-sources Qwen3.8-2.4T-A95B, first Max-tier weights release

Late on August 12, Alibaba Cloud ModelScope announced via WeChat that the Qwen team officially released the Qwen3.8-2.4T-A95B model weights. This marks the first time Qwen has open-sourced weights for a Max-tier flagship model; Qwen2.5-Max through Qwen3.8-Max were previously available only via Qwen Chat and DashScope API.

Qwen3.8-2.4T-A95B is a causal language model with 2.4T total parameters, using a sparse MoE architecture that activates 95B parameters per token. Each MoE layer contains 512 experts; each token selects 10 routed experts plus 1 shared expert. Context window natively supports 262,144 tokens and can be extended to 1,010,000 tokens, with multi-step MTP training applied.

Qwen3.8 continues the Qwen3.5 hybrid architecture, focusing on end-to-end task completion for coding, office work, scientific research, and long-horizon agent tasks. Official benchmarks span coding agent, general agent, professional work, and long-context dimensions; results vary against Opus 4.8, Fable 5, and GPT-5.6 Sol, with Qwen3.8 topping the listed models on PaperBench and IFBench.

The model can be deployed via SGLang, vLLM, and TokenSpeed inference engines. Open-source AI project Unsloth AI used dynamic 1-bit layered selective quantization to compress Qwen3.8-2.4T-A95B from 4.9TB to 397GB, a 91% reduction. With Unsloth-Desktop, machines with combined system memory and VRAM of 410GB or above can run the model locally.