Alibaba Previews Qwen3.8 Max and Promises to Give the Weights Away
Alibaba's Qwen team previewed a 2.4-trillion-parameter multimodal model it claims is second only to Claude Fable 5, with open weights promised but no date, licence or independent benchmarks yet.
Three days after Moonshot’s Kimi K3 landed, Alibaba answered. On July 19 its Qwen team previewed Qwen3.8-Max-Preview, which it describes as a 2.4-trillion-parameter model and, in its own words, “second only to Fable 5” among the systems it tested. It is the team’s first multimodal model above a trillion parameters, meaning one model that handles text, images, video and documents rather than text alone.
The published specifications are substantial. Integration metadata for Qwen Cloud lists a context window of 983,616 tokens and a maximum output of 131,072 tokens. A token is roughly a short word or word fragment, so that context window is around a million pieces of text held in mind at once, comfortably book-length. The preview is available now through Alibaba’s Token Plan, Qoder and QoderWork at ten percent of standard pricing, with Token Plan starting around six US dollars a month internationally and 39 yuan a month in China. The Qwen account also posted that the model will go open-weight, meaning anyone could download and run it, though no date and no licence have been named.
Two caveats deserve equal billing with the headline. First, the parameter count and the “second only to Fable 5” ranking are Alibaba’s own claims; no independent evaluation has been published. Second, the number of active parameters per token is undisclosed. That sounds like trivia, but it decides everything about cost: very large models today usually activate only a fraction of their parameters for any given request, so without that figure nobody can estimate what it would actually cost to run a 2.4-trillion-parameter model on your own hardware. An open-weight release from Alibaba’s flagship Max tier would also break the company’s own pattern, since Max models have so far stayed closed.
Step back and the shape of the week is clear. Chinese labs are no longer competing on price alone. Kimi K3 arrived at 2.8 trillion parameters with weights promised for July 27, Z.AI shipped GLM-5.2, and now Alibaba has a 2.4-trillion multimodal preview with an open-weight promise attached. Whether the models fully match Western frontier systems is still contested, but the release cadence and the willingness to publish weights are real and are putting pressure on the pricing of closed models.
What this means for you: For most people, nothing today. This is a preview with vendor-supplied numbers, and the sensible move is to wait for independent testing rather than a leaderboard claim. If you build with AI, the practical note is that another very large multimodal model may soon be downloadable, which is worth planning for even though the serving cost is currently unknowable. And if you simply want cheaper, better tools, competition of this kind is the mechanism that eventually delivers them.
Sources
The Director of District 9 Made a 13-Minute Film Entirely With AI, and Now Wants to Make a Feature
Neill Blomkamp released Nightborne, a sci-fi horror short generated with Seedance 2.0. He licensed the faces and voices of 32 real people, hired human concept artists, and says a full-length film is next.
