DeepSeek and Moonshot AI have trained their latest models largely on output from OpenAI’s GPT-6 Astra. The models in question are Kimi K3.1 Pro and DeepSeek V4.1-Flash 1017.
According to The Information, which cites people with direct knowledge of the matter, both Chinese AI developers relied heavily on the responses from the powerful OpenAI model. Moonshot AI released Kimi K3.1 Pro last week. DeepSeek launched the updated version, V4.1-Flash 1017, on October 17. Both models are said to have benefited from the distillation.
OpenAI, DeepSeek, and Moonshot AI did not immediately respond to requests for comment from Reuters. The news does not come entirely out of the blue. In a blog post dated September 30, OpenAI said it had linked a key component of a distillation campaign, which began in July, to individuals associated with Moonshot AI. The company subsequently blocked or restricted fraudulent accounts, tightened registration checks, and expanded network monitoring.
Catching up faster and more inexpensively
In distillation, a model learns from the responses of a larger, often proprietary model. This lets it gain comparable capabilities at a fraction of the training cost and computational power. Large-scale distillation via a commercial API may violate the terms of service. Companies such as Anthropic, OpenAI, and Google argue that competitors can use this method to replicate advanced capabilities more quickly and at a lower cost.
Earlier this year, OpenAI accused the Chinese lab of using accounts affiliated with DeepSeek to automatically collect responses for distillation.
Anthropic is also grappling with this phenomenon. The company says it has detected and disrupted campaigns by seven China-based labs since February. Anthropic’s new measures have reduced the amount of data stolen in distillation attacks by about 90 to 99 percent. Among other things, Anthropic now protects reasoning transcripts. In addition, Anthropic has permanently blocked tens of thousands of malicious accounts.