Alibaba's Qwen3.8 packs 2.4T parameters, claims second only to Claude Fable 5

For AI practitioners, the debut of Qwen3.8 at WAIC 2026 is both thrilling and frustrating. The model's massive 2.4 trillion parameter scale and native ability to handle images, video, and documents signal a serious technical leap. Yet the lack of any published benchmarks or model card—a stark contrast to the transparent launch of Qwen3.7-Max—leaves the community unable to assess the 'second only to Claude Fable 5' claim, raising fundamental questions about verification in an era of headline-grabbing AI announcements.

Alibaba Group Holding Ltd. sent a jolt through the AI community on July 19, 2026, with the surprise preview of Qwen3.8, its most powerful model to date, at the World Artificial Intelligence Conference in Shanghai. The company boldly claimed that the new system, offered as a “preview build” named Qwen3.8-Max-Preview, is second only to Anthropic’s Claude Fable 5 among frontier models. That assertion, if verified, would mark a dramatic leap for the Chinese tech giant and potentially reshape the global AI hierarchy. But the preview arrives with a conspicuous gap: Alibaba published no benchmark scores, no model card, and no comparison data against its own Qwen3.7-Max, which itself was launched in May with a full suite of evaluated metrics, including a 56.6 on the Artificial Analysis Intelligence Index. The silence around the new model’s performance raises immediate questions about transparency and whether the claim can be substantiated.

“

The company boldly claimed that the new system, offered as a “preview build” named Qwen3.8-Max-Preview, is second only to Anthropic’s Claude Fable 5 among frontier models.

Qwen3.8 represents a significant scaling milestone for Alibaba. At 2.4 trillion parameters, it is the first member of the Qwen family to cross the trillion-parameter threshold, and it also marks a shift toward native multimodality. The model can process images, video, and documents alongside text—capabilities that the earlier Qwen3.7-Max did not natively integrate. This mirrors a broader industry trend in which frontier labs are converging on multimodal architectures. Yet without disclosing the number of activated parameters—the actual subset of weights used during inference—the true computational footprint remains unclear. The lack of a model card, a document increasingly expected as a standard of responsible AI release, compounds the opacity. For developers and enterprise adopters, the absence of performance benchmarks handicaps any immediate evaluation of whether Qwen3.8 is worth the infrastructure investment, especially when priced at 10% of the standard rate during a trial period through Alibaba’s Token Plan and Qoder platforms.

The timing of the launch places Qwen3.8 squarely in a competitive domestic landscape. Just three days earlier, on July 16, Chinese AI startup Beijing Moonshot AI Technology Co. Ltd. released Kimi K3, a 2.8-trillion-parameter model that also pushes into the multi-trillion-parameter tier. Both launches underscore how Chinese developers are rapidly scaling to match the largest U.S. labs, a domain that until recently was dominated by a handful of Western companies. Alibaba, which has been positioning the Qwen series as a default AI backbone for Chinese enterprises and government, is likely motivated by the need to stay ahead of agile domestic rivals while simultaneously projecting parity with U.S. leaders. The claim of being “second only to Fable 5” is as much a marketing message as a technical one, designed to reassure both the domestic market and international observers that China remains a serious contender in the AI race despite ongoing export controls on advanced chips.