Alibaba Qwen 3.8-Max (2.4T) and 27B Open-Weights Models Released; Aims at Long-Horizon Coding and Agentic Work
Alibaba Qwen announced Qwen3.8-Max, a 2.4-trillion-parameter model described as its "most capable to date," with a commitment to release open-weights within one week. The model is also released as an API at $2 input/$6 output per million tokens, with $0.25/M for cached tokens. Simultaneously, Qwen is releasing Qwen3.8-27B as open-weights, covering both the large frontier tier and a practical open-weights tier. The models are positioned around coding, autonomous research, long-horizon planning, and native multimodal reasoning.
Qwen3.8-Max claims 95B active parameters per token (implying ~4% sparse activation in MoE), a 1M-token context window, and compatibility with OpenAI and Anthropic protocols. Benchmark claims include PaperBench 93.0, CoWorkBench 74.8, and WideSearch 81.9. Demonstrated capabilities span 10+ days of autonomous coding, 500+ optimization turns for chip design, and 365-day e-commerce strategy simulations. In third-party evaluations, the model debuted at #4 on the Frontend Code Arena and shows strong performance on vision and design-heavy tasks, with the model class comparable to other giant sparse open models like Kimi K3 and GLM-5.2.
The release signals renewed momentum in Alibaba's open-weights program following the Qwen exodus last year and management transitions. Qwen3.8-Max is available on Qwen Studio, API, Command Code, and partner platforms including Baseten, Hermes Agent, and Venice. Availability across both closed API and forthcoming open-weights versions creates dual-path deployment options.
For architecture evaluators, Qwen3.8-Max's open-weights roadmap and multimodal agentic design represent a competitive directional shift in the frontier open-weights tier. The model's 1M context and sparse activation approach sit in the same deployment class as Kimi K3, suggesting the Chinese open-weight frontier is now competing directly with Western closed models in coding and agentic workflows.