Alibaba released Qwen3.8-Max on August 3, 2026—a 2.4-trillion-parameter sparse MoE with 95B active parameters, native text+image+video input, and a 1M-token context window. The flagship API went live on QwenCloud ($2/$6 per million input/output tokens, matching GPT-5.6 Sol pricing) and is available through Vercel AI Gateway and OpenRouter. Critically, Alibaba committed to releasing open weights for both Qwen3.8-Max and a new Qwen3.8-27B during the week of August 10—the first time Alibaba has ever open-sourced a Max-class model, and without a disclosed license yet.
On agentic tasks, Qwen3.8-Max leads proprietary frontiers: it scores 86.1 on OSWorld-Verified (beating GPT-5.6 Sol Max at 83.2 and Claude Fable 5 at 85.0), 93.0 on PaperBench, and 86.6 on Terminal-Bench 2.1. Coding is tighter: it posts 67.7 on SWE-bench Pro (behind Fable 5's 80.0 and Opus 4.8's 69.2), but the gap is narrower than prior Qwen generations. On GPQA Diamond (science reasoning), it ties Fable 5 at 92.6. Arena placements show #4 in Frontend Code (1,668 Elo) and #2 in Vision (1,305), confirming real-world competitive performance. The 27B variant specs are unconfirmed but benchmarking hints from third-party tests suggest 70B-class capability on agentic workloads.
For architects: Qwen3.8-Max's open-weight commitment is the structural break. If both Max and 27B land on Hugging Face this week with Apache 2.0 (or equivalent) licensing, you acquire a 2.4T flagship and a practical 27B local model in the same release, eliminating the open-source-vs-frontier-capability tradeoff that has defined 2026. The $2/$6 API parity with proprietary frontiers plus open weights resets the cost structure for self-hosted agent fleets. Watch the license text on drop day—that determines deployment freedom.