Alibaba’s giant Qwen model raises stakes in China’s AI race

A launch measured in scale and ambition

Alibaba has unveiled Qwen3.8-Max, a mixture-of-experts model with 2.4 trillion total parameters and 95 billion activated for each query. The company says the system can process as many as 1 million tokens at once, spanning text, images and video, while tackling coding, research and other extended tasks.the-decoder +1 Its weights are due to be released next week, with access already available through QwenCloud.the-decoder

Investors treated the announcement as evidence that Alibaba remains a contender at the frontier. Its Hong Kong-listed shares closed Monday up 7% at HK$125.20, while the New York-listed stock gained 4.5%.caixinglobal +1 The reaction also reflects the commercial pitch: Alibaba paired the model with a public beta of QwenWork, an enterprise agent intended to bring the technology into workplace tasks.caixinglobal

Long-running tasks become the selling point

Rather than emphasizing short chatbot exchanges, Alibaba is promoting work that unfolds over days. In one company-run demonstration, Qwen3.8-Max spent 16 days developing a command-line coding tool, producing 265 commits and 127 pull requests without human intervention.the-decoder Another test had it reproduce a research paper, run 33 GPU training jobs and then improve the paper’s result on a math benchmark.the-decoder

The model’s sparse architecture matters because its headline parameter count is not the same as the computing load for every response. Activating 95 billion parameters at a time is designed to contain cost and latency, even as the full network carries far more capacity.sbs That makes efficiency, not scale alone, central to whether the release can win sustained developer and enterprise use.

Benchmark claims still need outside scrutiny

Alibaba says its model posts results comparable to, or better than, Anthropic’s Fable 5 on some tests, and Bloomberg described the release as the company’s biggest model yet.bloomberg CNBC reported that it ranked second in Vision Arena and fifth in Text Arena, but those comparisons were shared by Alibaba.cnbc Independent verification remains pending, and parameter counts do not automatically translate into better performance.the-decoder +1

Competition is intensifying inside China as well as against US laboratories. Moonshot AI’s recently released Kimi K3 has 2.8 trillion parameters and the same million-token context length, while independent tests cited by The Decoder found it lagged leading Western systems in cyber capabilities and complex mathematics.the-decoder Qwen3.8-Max therefore raises expectations, but its open-weight release and real-world deployment will provide the more consequential test of Alibaba’s claims.