Anthropic accuses three Chinese AI companies of ‘extensive distillation’

“We discovered an industrial-scale campaign by three AI laboratories including DeepSeek, Moonshot and MiniMax, which illegally exploited Claude’s capabilities to improve their own AI models,” Anthropic said on its blog on February 23. “This group conducted more than 16 million transactions with Claude through approximately 24,000 fraudulent accounts, violating our terms of service and regional access restrictions.”

In the post, the US AI company said it “always supports export controls to maintain America’s leadership in the field of artificial intelligence”, while attacking “distillation” to weaken controls. The campaign carried out by the three Chinese companies is “increasing in intensity and sophistication”, with a very short duration of action, and a threat “beyond the scope of any one company or region”.

“Addressing this problem requires quick and coordinated action between industry actors, policymakers, and the global AI community,” Anthropic urged, asserting that the attack copied from DeepSeek, Moonshot, and MiniMax violates the company’s terms of service. However, it is unclear whether this violates international law or not.

DeepSeek, Moonshot and MiniMax have not yet commented.

In the world of AI, the concept of “distillation” refers to the “transfer of knowledge” from one model to another in a teacher-student fashion. “Distillation is a technique designed to transfer the knowledge of a large pre-trained model (teacher) into a smaller model (student), allowing the student model to achieve performance equivalent to the teacher model,” scientists Vishal Yadav and Nikhil Pandey told the publication. Forbes. “This technique helps users leverage the quality of large language models (LLMs), while reducing inference costs.”

Leading AI models from OpenAI, Google, Meta or Anthropic essentially learn on their own from scratch with huge amounts of raw data – a process that often takes months and costs tens of millions of dollars or more. However, when another company comes along and takes advantage of the results already from these leading AIs, the “distillation” process can help create a good model in weeks, even days, at significantly less cost.

Early last year, told FTOpenAI said it detected signs of “distillation” from DeepSeek. However, according to Mashablemany people at that time did not sympathize but turned to ridicule because OpenAI and other large AI companies believed that they had “absolute rights” to train their models on copyrighted works without asking permission or paying fees.

Earlier this month, Google’s Threat and Intelligence group (GTIG) also “detected an increase in widespread pattern extraction attempts or distillation attacks.” However, the company did not name any specific names.

DeepSeek became famous early last year when it launched a high-performance AI model with limited costs. The company claims it used only 2,048 Nvidia H800 graphics cards and $5.6 million to train the V3 model with 671 billion parameters, a fraction of the amount OpenAI and Google had to spend to train a model of equivalent scale.

Moonshot AI is China’s new unicorn, famous for its Kimi K2. This open source model allows developers to skip training from the beginning and freely use, tweak, and expand as they wish. With a large number of parameters, Kimi K2 outperformed competitors ChatGPT 4.1 in the coding ability test and Claude 4 Opus in the scientific knowledge test. Kimi K2 ranks 9th on our open source model rankings Hugging Face as of August last year.

MiniMax is also a prominent Chinese AI startup, belonging to the “AI Tigers” group (tigers leading the LLM wave). According to Reutersin its product ecosystem, MiniMax builds many multimodal AI platforms that can process text, voice, images and videos, serving hundreds of millions of domestic users.