Alibaba’s Qwen team on Sunday unveiled Qwen 3.8, a 2.4-trillion-parameter multimodal model the company says matches frontier systems and trails only Anthropic’s Fable 5. A preview is live now through Alibaba’s Token Plan subscription, its Qoder coding platform and the QoderWork productivity suite — at 10 percent of standard pricing. The full model is promised as an open-weight release “soon.”
That last word is the story. Alibaba’s Max-tier flagships have always stayed closed: the open Qwen catalog built one of the world’s largest developer bases, while the Max models made the money. Qwen 3.8 breaks that split. It’s the team’s first multimodal model above a trillion parameters — it processes images, video and documents — and Alibaba says it should beat Qwen 3.7-Max most clearly in coding and complex productivity work: full-stack development, data analysis, office workflows. Giving that away is not a research decision. It’s a pricing weapon.
The Kimi K3 subtext
The target is hard to miss. Moonshot’s Kimi K3 spent the past week as China’s open-model champion, with benchmarks pressing the frontier — but Moonshot has routed actual access through its own app and API while the promised weights wait, converting attention into revenue. It worked: $300 million in annual recurring revenue as of June, and Bloomberg reported Sunday the company plans an IPO in as little as six months. An open-weight 2.4T Qwen lands squarely on that business model — why pay Moonshot’s meter if Alibaba’s biggest model is free to download?
The timing compounds it. The drop caps China’s loudest AI week — the same WAIC weekend that produced Ant’s catalog of 200 prebuilt AI “digital employees” and a parade of humanoid robots — and it arrives hours before Anthropic’s free Fable 5 access expires for millions of paying subscribers.
What’s missing
Plenty. No benchmark results have been published. No license, no release date for the weights. And Alibaba hasn’t disclosed how many of the 2.4 trillion parameters are active at inference — the number that actually determines cost and speed in sparse mixture-of-experts models. Until the weights and the scores exist, “second only to Fable 5” is a marketing sentence, not a measurement.
Our take: Open weights have become China’s competitive weapon — aimed less at Silicon Valley than at each other. Alibaba isn’t donating its crown jewel; it’s spending it, to blunt a rival’s IPO run. For everyone else the math rhymes with DeepSeek’s moment: frontier-adjacent capability at collapsing prices, again. The winners are buyers of intelligence. The losers are anyone whose business depends on charging rent for a closed model that’s only slightly better than a free one.
What to watch
- The weights, the license, the benchmarks. A “soon” that slips — or a restrictive license — would say the claim outran the model.
- Moonshot’s counter. K3 price cuts, an accelerated weights release, or a faster IPO clock all signal how much this stung.
- The US answer. Anthropic’s paid Fable 5 era starts Monday; every open-weight release at this scale makes closed-model pricing harder to defend.