Source: MarkTechPost — 2026-07-19
Summary
Alibaba previewed Qwen3.8-Max at the World AI Conference in Shanghai, a 2.4-trillion-parameter sparse mixture-of-experts model handling text, images, video, and documents with a 1M-token context window — its first multimodal model above 1 trillion parameters. The preview shipped days after Moonshot's open-weight Kimi K3 launch, with Alibaba claiming performance "second only to Fable 5," but without any accompanying benchmark results, model card, or license terms.
Key Takeaways
- Qwen3.8-Max-Preview is a sparse MoE architecture at 2.4T total parameters, Alibaba's largest multimodal model to date, inheriting the 1M-token context window from Qwen 3.7 Max.
- It's available now only through Alibaba's own hosted channels — Token Plan, Qoder, and QoderWork — priced at 10% of standard rates, not as downloadable open weights (open weights are promised "coming soon," per Alibaba, but not yet delivered).
- The "second only to Fable 5" performance claim currently has no independent or Alibaba-published benchmark numbers behind it — no model card, no published eval table, no license — making this a preview announcement to watch rather than a verified result.
- Timing is the real story: this landed just two days after Moonshot's Kimi K3 open-weight release, underscoring how fast Chinese labs are cycling through frontier-scale announcements this cycle — worth revisiting once actual benchmarks or open weights ship.