Qwen3.8-Max is 📝Alibaba's flagship multimodal large language model and the largest in the 📝Qwen family, a 2.4-trillion-parameter sparse mixture-of-experts system launched on August 3, 2026.
The model belongs to the 📝Qwen3.8 generation and activates roughly 95 billion parameters per token, combining sparse mixture-of-experts routing with hybrid attention on the Qwen3.5 foundation. Its API offers a one-million-token 📝context window (about 991,000 input and 131,000 output tokens), accepts text, image, and video input, and launched at $2 per million input tokens, $6 per million output tokens, and $0.25 per million cached input tokens.
At launch Alibaba reported fifth place in Text Arena, second in Vision Arena, and fourth in Frontend Code Arena. Reported scores include 92.6 on GPQA Diamond, 86.6 on Terminal-Bench 2.1, 86.1 on OSWorld-Verified, and 73.5 on FrontierSWE, up from 40.7 for its predecessor; independent verification remained limited in the weeks after release. Alibaba says the model ran a 16-day unattended software project that produced oh-my-cli, an open-sourced self-evolving agent framework, and introduced RecreationBench, in which the model rebuilds applications from scratch by interacting with them live, without source code or internet access.
Qwen3.8-Max is available through Alibaba Cloud Model Studio and on QwenWork, Alibaba's workplace AI agent platform. Open weights followed on August 12 as Qwen3.8-2.4T-A95B, a text-only checkpoint with a 262,144-token native context, so vision input and the full million-token window remain features of the hosted API.