Alibaba released Qwen3.8-Max on August 3, 2026, a 2.4-trillion-parameter sparse mixture-of-experts model with 95 billion parameters active per token, 1 million token context window, and multimodal input (text, image, video). The model is available via API on QwenCloud and Alibaba Cloud Model Studio at $2 per million input tokens, $6 per million output, and $0.25 cached input. Alibaba committed to releasing open-source weights within one week, making Qwen3.8-Max the first Max-class model Alibaba has ever open-sourced. A smaller Qwen3.8-27B is also going open.
On Arena.AI benchmarks at launch, Qwen3.8-Max ranked fifth on Text Arena and second on Vision Arena globally. Alibaba's own benchmark scorecard shows 86.6 on Terminal-Bench 2.1 (ahead of Claude Opus 4.8 at 84.6 but behind GPT-5.6 Sol at 88.8), 92.6 on GPQA Diamond, and 67.7 on SWE-bench Pro (versus Claude Fable 5's 80.0). Major gains are in multimodal and agentic reasoning, not pure reasoning. Alibaba claims Qwen3.8-Max trails only Claude Fable 5 overall, though third-party verification (Artificial Analysis, LMArena) had not scored it as of early August.
Qwen3.8-Max demonstrates extended agentic reasoning: Alibaba showed it autonomously coding and iterating a self-evolving software harness over 10+ days with user feedback, and reproducing a machine learning research paper end-to-end with 33 GPU training rounds (125 hours wall-clock) and 7,600 lines of generated code. When entered into a live coding contest with 526 human teams, it beat 87% of the field. Multimodal handling includes summarizing 100-page documents or 100-hour livestreams into searchable knowledge bases.
For architects: API pricing of $2/$6 (versus Claude Opus 5 at $30, GPT-5.6 Sol at $35) makes this aggressively underpriced for agentic/multimodal work, but self-hosting the 2.4T-parameter full model when weights ship will require multi-node datacenter deployment. The 27B open-weight checkpoint is the realistic on-premise path. Third-party benchmarking is still missing; Alibaba's self-reported benchmarks (Terminal-Bench, GPQA) lack independent corroboration as of early August. Assume parity with Fable 5 for planning purposes until Artificial Analysis or LMArena publish scores.