Alibaba’s newest AI model can code alone for 16 days straight. It beats GPT-5.6 Sol and Claude Fable 5 on seven separate benchmarks. Alibaba open-sourced the weights anyway — inviting the world to check the claims rather than take them on trust.
Alibaba released Qwen3.8-Max on Monday, its largest and most capable AI model to date. The 2.4-trillion-parameter system delivers “comprehensive improvements across coding, work, research, and long-horizon tasks,” according to the Qwen team’s own announcement. One internal test had the model working alone for 125 continuous hours — roughly five days — rebuilding a math research paper’s experiment from scratch and then improving on it. In a separate contest, Qwen3.8-Max competed against 526 human teams under a 24-hour deadline and beat all but 68 of them. Alibaba’s own benchmarks show the model beating both GPT-5.6 Sol and Claude Fable 5 across seven separate evaluations.
What’s Happening & Why It Matters
Annexing Anthropic’s Marketing
Alibaba’s stance isn’t accidental. The claim that Qwen3.8-Max can work unsupervised for days on end echoes almost how Anthropic has described Claude Fable 5 since its release. As TF covered in Anthropic’s Fable 5 Full Restoration, Fable 5 was placed under US export controls because of concerns over its advanced capabilities. Alibaba is claiming performance that matches, and in some benchmarks beats, the very model Washington judged dangerous enough to restrict.
Anthropic’s newest model, Claude Opus 5, launched 24 July and is built to run unsupervised for long stretches too. Anthropic says the model handles unfamiliar problems faster than any prior version. Both companies are making nearly identical claims about the same capability, within the same two-week window.
Open Weights: A Strategy of Trust
Alibaba’s most striking decision isn’t the benchmark scores. It’s the choice to open-source Qwen3.8-Max’s weights on Hugging Face and ModelScope, letting any researcher verify the company’s claims rather than accepting them on faith. Anthropic and OpenAI keep their most capable models closed, accessible through paid APIs. Alibaba’s answer to scepticism about Chinese AI progress is radical transparency instead of marketing.
Qwen3.8-Max activates 95 billion of its 2.4 trillion total parameters per request — a mixture-of-experts design that keeps inference costs down while preserving the model’s full capacity for complex tasks. The model supports a context window up to 1 million tokens and processes text, images, and video, putting it on functional par with the multimodal capabilities OpenAI and Anthropic have built into their own top-tier models.
Other Fast-Acting Chinese Labs
Qwen3.8-Max arrives in a crowded field of recent Chinese releases. Moonshot AI’s Kimi K3 launched with 2.8 trillion parameters as an open-weight model competing with US frontier offerings, as TF covered in Kimi K3 Halts New Signups as Demand for the Chinese AI Model Skyrockets. DeepSeek released DeepSeek-V4-Flash the same week, delivering major coding and agentic improvements at a fraction of US model pricing. Three separate Chinese labs shipped competitive frontier models within a month of each other.
That pace connects to the tension TF has tracked in Washington throughout 2026. Silicon Valley executives and federal lawmakers have increasingly treated the AI race with China as an existential concern rather than a routine commercial rivalry. Anthropic itself accused operators linked to Alibaba of running approximately 29 million fake exchanges with Claude in a distillation attack, warning that such campaigns “turn hundreds of billions of dollars in American investment and R&D into a massive subsidy for our geopolitical competitors.” Alibaba’s own shares jumped as much as 7.3% on the Qwen3.8-Max release, their best performance in a month.
TF Summary: What’s Next
Qwen3.8-Max’s full model weights are scheduled for open release on Hugging Face and ModelScope next week. Alibaba continues expanding access through its Token Plan subscription and Qoder agentic platforms. No independent third-party benchmark verification has yet confirmed Alibaba’s specific performance claims against Fable 5 and GPT-5.6 Sol. Anthropic’s distillation attack allegations against Alibaba are under Congressional review.
MY FORECAST: Expect independent benchmark verification within two weeks of the open-weight release, given how Alibaba invited scrutiny by publishing the weights rather than keeping the model closed. If Qwen3.8-Max’s real-world performance holds up against Alibaba’s own claims, expect US lawmakers to cite the release in upcoming hearings on AI export controls and distillation risk. The genuine story here isn’t whether Alibaba caught up to Anthropic on a single benchmark. It’s that three separate Chinese labs ship frontier-competitive models on a monthly cadence, and open weights make every one of those claims checkable by anyone with the compute to run them.
Related Stories
- Anthropic’s Fable 5 Full Restoration
- Kimi K3 Halts New Signups as Demand for the Chinese AI Model Skyrockets
- Anthropic Accuses Alibaba of 28.8 Million Fake Queries to Clone Claude