Alibaba Cloud used its home turf at the World Artificial Intelligence Conference in Shanghai to preview Qwen3.8-Max, the newest flagship in its Qwen lineup.
The company built the model with 2.4 trillion parameters and claimed it trails only Anthropic’s Fable 5 among frontier models.
That’s a striking claim to make in a single Sunday X post, and it arrives with almost nothing to independently verify it.
What Alibaba Actually Confirmed, Versus What It Only Claimed
The confirmed details are fairly specific. Qwen3.8-Max is a sparse Mixture-of-Experts model that supports text, images, video, and documents, features a 1-million-token context window, and becomes the first Qwen model above one trillion parameters to support multimodal input.
Alibaba has already made the preview available through its Token Plan, Qoder, and QoderWork platforms at 10% of standard pricing. The company also supports major coding tools, including Claude Code, Cursor, Codex, Cline, and its own Qwen Code.
Everything beyond those details becomes far less certain. Alibaba published no benchmark table, model card, or technical report to support its “second only to Fable 5” claim. Multiple outlets have described that ranking as Alibaba’s own assessment rather than an independently verified benchmark.
The company also hasn’t disclosed the model’s active-parameter count, an important omission for a Mixture-of-Experts architecture because total parameters reveal little about how many parameters each query actually activates or how expensive the model is to serve.
Alibaba also provided no comparison with Qwen3.7-Max, which it released just two months ago in May, instead describing Qwen3.8-Max only as “continuously evolving.”
The company promised to release open weights “soon” but attached neither a launch date nor licensing terms. Alibaba made the same promise for Qwen3.7-Max without ultimately releasing the weights.
Why the Timing and the Missing Math Both Matter
Qwen3.8-Max arrived just two days after Moonshot AI launched Kimi K3, a 2.8-trillion-parameter open-weight model that had already generated significant momentum. Most independent coverage has interpreted Alibaba’s timing as a direct response rather than a coincidence.
The bigger question is practical rather than competitive. Independent analysts estimate that a 2.4-trillion-parameter model running at 4-bit precision requires roughly 1.2 terabytes of memory just to store its weights.
That exceeds the combined memory available across eight Nvidia H200 GPUs, each equipped with 141GB.
That calculation shifts the focus away from leaderboard rankings and toward deployment. Alibaba now needs to demonstrate whether it can deliver a practical version through a smaller active-parameter configuration, an efficient quantized checkpoint, or a distilled model.
As described today, the full model remains difficult to deploy at meaningful scale.
As of this writing, neither Artificial Analysis nor LMArena has independently evaluated Qwen3.8-Max. Until those results appear, Alibaba’s claim that the model ranks “second only to Fable 5” rests entirely on the company’s own internal testing.
None of that makes Qwen3.8-Max insignificant. A 2.4-trillion-parameter multimodal model from one of China’s largest cloud providers represents an important milestone in the rapidly evolving open-weight AI race.
However, Alibaba still needs to publish benchmark results, disclose the model’s active-parameter count, and compare it directly with Qwen3.7-Max before the broader AI community can independently verify its frontier-level claims.
Source: FoneArena, "Alibaba Unveils Qwen3.8-Max Preview with 2.4T Parameters; Open Weights Coming Soon"
