Alibaba Previews 2.4-Trillion-Parameter Qwen3.8, With Weights Promised but No Benchmarks Shown

Alibaba has previewed Qwen3.8-Max, a 2.4-trillion-parameter multimodal model it describes as one of the most capable systems available and, by its own account, second only to Claude Fable 5 among the models it benchmarked against. The Russian-language AI channel AI ML Big Data reported the preview, noting that it is testable through Alibaba's token service and that the company positions it alongside leading frontier models. It is the Qwen team's first multimodal model above one trillion parameters, able to process images, video, and documents.

The timing is competitive rather than coincidental. The preview lands days after Moonshot's Kimi K3, and the open-weight pledge directly contests the downloadable tier that Chinese labs have made their own. The substance, however, is thin. As of the July 19 announcement there was no model card on Hugging Face, no license, no release date, and no disclosed active-parameter count for what is presumably a sparse mixture-of-experts design.

For a mixture-of-experts (MoE) model, the active-parameter figure is the number that governs serving cost, and its absence makes the 2.4-trillion headline close to meaningless for capacity planning. No published benchmark numbers accompany the claim that the model is second only to Fable 5, which is a vendor assertion with no independent reproduction. The verifiable facts today are the parameter count, the multimodal capability, and the stated intent to open the weights.