Alibaba Previews Qwen3.8 Max and Aims Straight at Fable 5

Alibaba's Qwen team pulled the wrapper off its biggest model yet on July 19, and the headline number is hard to ignore. Qwen3.8 Max carries 2.4 trillion parameters, which puts it in a tier only a few frontier labs have reached. The team says it lands second only to Anthropic's Fable 5 across the benchmarks it ran, a claim that would place a Chinese model within touching distance of the current Mythos-class leader.

What is actually new

Qwen3.8 Max is built on a sparse mixture-of-experts design, so only a fraction of those 2.4 trillion parameters fire on any given request. That is how a model this size stays affordable to run. The more meaningful shift is that this is the first Qwen model above a trillion parameters to handle images, video, and documents rather than text alone. Earlier Qwen flagships were strong but text-bound. This one is properly multimodal.

For now it is a preview rather than a full release. The model is live through Alibaba's Token Plan, Qoder, and QoderWork at roughly a tenth of standard pricing, and the company has promised open weights "soon" without naming a date. The benchmark table, the model card, and the license are all still to come.

The timing is the story

The preview landed only days after Moonshot released Kimi K3, a 2.8-trillion-parameter open-weight model that briefly took the top coding spot. Put the two together and a pattern is obvious. China's leading labs are now shipping multi-trillion-parameter systems within days of one another, and both are pointing their marketing directly at the American frontier.

Read the benchmark claim carefully

The "second only to Fable 5" line rests entirely on Alibaba's own internal evaluations. No independent lab has tested Qwen3.8 Max, and the company has not yet published the numbers behind the claim. That is worth holding in mind. Self-reported benchmarks have a long history of flattering the model that ran them, and headline scores are among the easiest things in AI to game. We looked at why benchmark scores rarely tell you the best model earlier this month, and the caution applies here in full.

None of that makes the preview unimportant. A 2.4-trillion-parameter multimodal model from Alibaba, priced this aggressively, is a real marker of how fast the gap is closing. It just means the "second only to Fable 5" banner is a hypothesis for now, not a result. The open-weight drop, whenever it comes, is when the rest of us get to check.