Alibaba previews Qwen 3.8 (2.4T params) without benchmarks; claims second only to Fable 5

Alibaba's Qwen team previewed Qwen 3.8-Max-Preview on July 19 at the World AI Conference in Shanghai, a 2.4 trillion-parameter multimodal model positioned as "second only to Fable 5" among systems Alibaba benchmarked. The model is available through Token Plan subscriptions ($6 Lite tier, $68 Pro tier) at 10% of standard pricing, with open-weight release promised "soon" but no date or license disclosed.

Critically, Alibaba has published no independent third-party benchmarks to verify its performance claims—no model card, no technical report, no active-parameter count for the sparse mixture-of-experts architecture. The preview ships with 1M token context window, multimodal capabilities (text, images, video, documents), and enables reasoning with configurable intensity. The timing comes days after Moonshot AI's Kimi K3 (2.8T params) launched open-weight, signaling intense competition between Chinese labs for frontier-scale positioning.

For architects: Qwen 3.8 represents the third major Chinese model claim at or near the frontier in as many months (GLM 5.2, Kimi K3, now Qwen 3.8), but the lack of published benchmarks and unresolved tokenomics (serving cost depends on active-parameter count, still undisclosed) make production deployment premature. Test on your own workloads; benchmark claims from internal evals alone don't predict real-world performance until independent verification arrives.