Qwen3.8-Max

Qwen3.8-Max, released August 3, 2026, was the largest model Alibaba had built: 2.4 trillion total parameters with 95B active, a 1M-token context window, and hybrid thinking. Alibaba previewed it on July 19 and shipped it to the Alibaba Cloud API two weeks later, placing it fifth on Text Arena and second on Vision Arena at launch.

The launch numbers pointed at agents rather than raw coding. It scored 93.0% on PaperBench and 86.1% on OSWorld-Verified — both ahead of every model Alibaba benchmarked against, including GPT-5.6 Sol and Claude Fable 5 — along with 86.6% on Terminal-Bench 2.1 and 53.4% on JobBench. On conventional software engineering it was further back, at 67.7% on SWE-Bench Pro against the 80.3% Claude Fable 5 had posted at its own release two months earlier. Multimodal results were strong without a code tool: 88.4% on CharXiv chart reasoning and 82.0% on BabyVision.

Alibaba announced at release that it intended to publish the weights, and on August 12, 2026 it opened the text-only base model underneath the product, Qwen3.8-2.4T-A95B, under a custom licence — by a wide margin the largest weights any lab had published to that point, though the hosted Max kept vision input, non-thinking mode and the full 1M context for itself. Opening a Max-tier flagship cut against the pattern the Max line had followed since Qwen3-Max, and the smaller Apache-licensed Qwen3.8-27B followed two days later. The launch landed two and a half weeks after Moonshot AI's Kimi K3, in a stretch where the Chinese labs were shipping trillion-parameter models within weeks of each other. A month later Alibaba re-post-trained the same model and shipped it as Qwen3.8-Max-0902.