Alibaba's Qwen 3.8-Max Claims Anthropic-Level Performance in New China AI Wave

Alibaba Launches Qwen 3.8-Max, Its Largest AI Model Yet

Alibaba Group has released its largest artificial intelligence model to date, claiming performance on par with Anthropic PBC's flagship systems and adding a new chapter to China's accelerating challenge to US AI leadership.

The new model, Qwen 3.8-Max, has 2.4 trillion parameters. Results shared by Alibaba show it ranking above Moonshot AI's widely watched Kimi K3 on some benchmarks, and scoring roughly on par with — occasionally ahead of — Anthropic's Fable 5, a frontier model the United States briefly placed on an export control list because of its advanced capabilities. The launch lands in a crowded month for Chinese AI: Moonshot's Kimi K3 debuted last month, DeepSeek opened the API public beta of its V4 Flash flagship last week, and ByteDance and MiniMax each released video-generation models on Friday.

Markets responded immediately. Alibaba's Hong Kong-listed shares rose as much as 7.3% on Monday, the biggest one-day gain in nearly a month. Ling Vey-Sern, a managing director at Union Bancaire Privée, said many investors underestimate Chinese AI models because of US chip restrictions and general skepticism. 'The gap is much smaller than many believe and it is closing fast,' he said. 'Kimi K3, and now Alibaba's Qwen 3.8, are further proof.'

Alibaba said it will publish Qwen 3.8-Max's model weights for public download next week, letting users customize the technology — the latest twist in an increasingly fierce open-weight rivalry among Chinese developers. The company also set sharply competitive prices: $2 per million input tokens and $6 per million output tokens, well below leading US models even after accounting for token-efficiency differences. Alibaba says the model excels at autonomous coding and long-horizon execution, and that in internal testing it independently completed a 16-day software engineering project. A mixture-of-experts design, which activates only part of the parameters per request, is meant to cut compute costs and latency.

Where Qwen 3.8-Max Leaves the China–US AI Race

Self-reported benchmarks: treat the parity claim with caution

Alibaba's headline result — matching or beating Anthropic's Fable 5 on some tests — comes from the company's own evaluation, and no independent benchmark has yet confirmed it. That alone warrants skepticism about the word 'parity'. Yet the same caveat applied to Moonshot AI's Kimi K3 a month ago, and the pattern across several Chinese labs is harder to wave away: Kimi K3 impressed on constrained compute, DeepSeek shipped a low-cost flagship API, and Alibaba now fields its largest model yet. The rhythm of releases is itself evidence that China's ecosystem is iterating quickly, supporting the view expressed by Union Bancaire Privée's Ling Vey-Sern that the capability gap with US labs 'is much smaller than many believe and is closing fast'.

Open weights and low prices: Alibaba's two-pronged play

Alibaba is competing on both access and price. Releasing the model weights next week lets enterprises fine-tune and run Qwen 3.8-Max on their own infrastructure — the same play that built followings for DeepSeek and Moonshot. The $2-per-million input and $6-per-million output token pricing undercuts top US frontier APIs by a wide margin, and the mixture-of-experts design, which activates only part of the 2.4 trillion parameters per query, keeps serving costs down. For Alibaba's cloud arm — the company's growth engine — cheap, popular models are a funnel for developer traffic and, eventually, computing workloads.

The export-control shadow over a 2.4-trillion-parameter release

Anthropic's Fable 5 was temporarily placed on a US export control list, the clearest sign Washington already views frontier AI capability as a security matter. A 2.4-trillion-parameter open-weight model built by a Chinese company despite chip restrictions is exactly the kind of development that could invite a response — whether new limits on model weights, tighter chip rules, or restricted access for downstream users. At the same time, Chinese labs are proving they can reach near-frontier performance with limited compute, which undercuts the assumption that export controls alone preserve the US lead.

The domestic race is the real pressure cooker

Moonshot, DeepSeek, ByteDance and MiniMax have all released notable models within roughly the past month, and Alibaba's launch lands in the middle of that wave. The move reasserts Alibaba's position among China's strongest open-weight developers, but it also confirms that the release cycle has accelerated: frontier-scale launches are now arriving within weeks of each other, and each one drags prices down. Enterprises and developers are the immediate winners. The pressure falls on US labs whose premium pricing rests on a capability gap that is visibly shrinking.

What Investors and Developers Should Watch After the Qwen 3.8-Max Launch

For investors:

• Treat the benchmark parity claims as unverified until independent evaluations appear; the credibility test for Alibaba's numbers will likely come within weeks of the weight release.
• Watch the planned public release of Qwen 3.8-Max weights next week — it is a stronger signal for real adoption and Alibaba Cloud demand than the launch announcement itself.
• Track Alibaba's Hong Kong-listed shares beyond the 7.3% one-day jump on Monday; sustained momentum depends on third-party benchmark results and cloud revenue visibility, not the announcement spike.

For developers and enterprises:

• Evaluate Qwen 3.8-Max's $2-per-million input and $6-per-million output token pricing against US frontier APIs; even accounting for token-efficiency differences, the cost gap is large for high-volume workloads.
• Test the open-weight version once it ships — Alibaba's claims of strength in autonomous coding and long-horizon agents, including the 16-day internal engineering project, are directly verifiable on your own tasks.
• Given the near-weekly release pace from Moonshot, DeepSeek and ByteDance, benchmark competing Chinese models on the same workload before committing to one provider.