Alibaba's 2.4T Qwen3.8-Max lands second only to Fable 5

Alibaba previewed Qwen3.8-Max on July 19: 2.4 trillion parameters, its first multimodal model above 1T, handling text, images, video, and documents. The team says it sits second only to Anthropic's Fable 5 on the benchmarks it ran. Caveat worth noting — no benchmark table, model card, or license has shipped, and Alibaba won't say how many of those 2.4T parameters actually fire per token. It's a sparse MoE, so the headline number flatters. Open weights are promised "soon," no date.

It landed days after Moonshot's Kimi K3 (2.8T params, weights due July 27), which promptly became its own cautionary tale. Within 48 hours demand "pushed close to the limits of our current capacity" and Moonshot paused new subscriptions, redirecting spare GPUs to existing users. New slots reopen in controlled batches.

Meanwhile Google started surfacing original recipe links more prominently in AI Mode — a small concession in its widening standoff with publishers over AI search eating referral traffic.

Why it matters: Chinese labs have closed the capability gap faster than they've closed the compute gap — shipping frontier models they can't afford to serve.