Overview
alibaba qwen flagship multimodal sparse mixture-of-experts model announced July 19, 2026 during world-ai-conference. Disclosed size: 2.4 trillion parameters. Preview as Qwen3.8-Max-Preview on Token Plan, qoder, and QoderWork; open-weight release promised “soon.”
Claims vs Evidence
-
2026-08-03: qwen3-8-max GA — active params disclosed as ~95B in Alibaba Cloud Community blog; open weights next week (2026-08-03-alibaba-qwen38-max-alicloud-blog)
-
Company: “second only to claude-fable-5” among models it compared
-
Independent benchmarks (Artificial Analysis, LMArena, etc.): not published as of announcement day
-
Active-parameter count: preview-era undisclosed; Alibaba Cloud Community GA blog cites ~95B active for qwen3-8-max (still vendor-disclosed)
Warning
Ranking language is Alibaba’s self-claim. Do not treat as verified league standing until third-party evals land.
Scale Context
- Second-largest publicly disclosed Chinese model after kimi-k3 (2.8T, moonshot-ai)
- Continues Max-tier scale-up from Qwen3-Max-Preview (2025) and Qwen 3.6 Max Preview (Apr 2026)
Related
- qwen
- alibaba
- frontier-models
- open-weight-models
- mixture-of-experts
- claude-fable-5
- kimi-k3
- us-china-ai-competition
- qwen3-8-27b
- alibaba-cloud-model-studio