- Anthropic claims that Alibaba has illegally “absiphoned” knowledge from Claude through a series of distillation training runs and millions of prompts for downstream use.
- In a filing to the U.S. District Court for 10 June 2026, Anthropic alleges that Alibaba has scaled up distillation training runs on massive blocks of data, including smaller problems that could be used to train Claude.
- Anthropic says it has details on 25,000 shared accounts and 28.8 million exchanges, which it claims Alibaba used to train the Qwen3.8-Max model, and it says the “open weights” feature is being used to train it.
For more on this, see The Verge and vox.com.
Anthropic says that the Chinese company Alibaba has illegally “absiphoned” copyrighted data from Claude. According to the company, for this purpose it used millions of distillation training runs for downstream use of Claude, and then millions of prompts and downstream use of LLM Alibaba.
In its filing to the U.S. District Court, dated 10 June 2026, Anthropic claims that Alibaba scaled up distillation, using it to train Claude, for “agentic reasoning,” software engineering, and long-horizon tasks. The company also says it is not clear whether these are illegal, and that it is unclear whether this is a “zbirati” AI-zdatnist with the company’s models and whether it is not a “zbirati” AI-zdatnist with the company’s models.
Anthropic also claims that the company used the acquisition of information from U.S. AI laboratories, including restrictions on what is allowed for PRC AI laboratories, and that it used it for downstream use through distillation.
Anthropic also says it has evidence that the illegal distillation was used. According to the company, its training runs included 25,000 shared blocks with 28.8 million exchanges; it also says that the process continued from 10 June 2026.
In parallel with this, Alibaba’s distillation of the Qwen3.8-Max model is being carried out, which the company says is the best and most stable way to do so. According to The Verge, Alibaba says it is using it to improve the model’s performance through blog posts, and also says that it is currently planning to publish the weights (weights) of Qwen3.8-Max. The company also says that Alibaba is disclosing its model’s product characteristics and performance in benchmark tests, including Arena.AI, and that the model has 2.4 trillion parameters.