Alibaba Qwen Open Sources Qwen3.8-2.4T-A95B Flagship Model, 2.4 Trillion Parameters, 91% Size Reduction

Alibaba's Qwen team has open-sourced the weights of the Qwen3.8-2.4T-A95B model. The model has 2.4 trillion total parameters, with 95 billion parameters activated per token. It natively supports a context length of 262,144 tokens, expandable to 1.01 million tokens. Built on the Qwen3.5 architecture, the model delivers enhanced performance in programming, office productivity, scientific research, and long-cycle agent tasks. This marks the first time Qwen has made Max-level flagship model weights openly available.

Model deployment supports the SGLang, vLLM, and TokenSpeed inference engines, requiring the latest Recipe from each framework. The model operates by default in thinking mode, which cannot be disabled. The open-source AI project Unsloth AI applied dynamic 1-bit hierarchical selective quantization technology to compress the model size from 4.9TB to 397GB, achieving a 91% reduction in storage footprint. Using the Unsloth-Desktop tool, the model can run locally on devices with a combined memory and video memory of at least 410GB.

In benchmark tests, Qwen3.8-Max scored 93.0 points on PaperBench, surpassing GPT-5.6 Sol, Fable 5, and Claude Opus 4.8. It achieved the top score of 86.1 points on OSWorld-Verified and 91.5 points on the parameterized CAD benchmark, outperforming Fable 5 and others. However, it scored 86.6 points on TerminalBench 2.1, below GPT-5.6 Sol's 88.8 points; 67.7 points on SWE-bench Pro, trailing Fable 5's 80.0 points; and 68.3 points on VideoMME v2, below GPT-5.6 Sol's 71.1 points.

The Qwen team tested the model's autonomous task-completion capabilities, including continuous autonomous programming for approximately 16 days, independent reproduction of research papers, participation in algorithm competitions, and operation of a virtual e-commerce enterprise. During the autonomous programming task, the model built a self-evolving Harness to perform requirement collection, code generation, and bug fixing.

Regarding API pricing, Qwen3.8-Max charges RMB 12 per million input tokens and RMB 36 per million output tokens domestically, with implicit cache hits at RMB 1.5 per million tokens. Overseas pricing is USD 2 per million input tokens and USD 6 per million output tokens, with implicit cache hits at USD 0.25 per million tokens.