Technology media outlet The Information reported on the 28th, citing people directly familiar with the matter, that Chinese AI developers DeepSeek and Moonshot AI used large amounts of output from OpenAI’s GPT-6 Astra model to train their latest models. Reuters said in its account of the report that OpenAI, DeepSeek, and Moonshot AI did not immediately respond to requests for comment.
This practice is known as “distillation”: repeatedly querying a more capable model, collecting its answers, and then using those outputs to train another model. As U.S.-China tech competition intensifies, distillation has become a contentious issue among major AI companies.
Moonshot AI's Kimi K3.1 Pro and DeepSeek V4.1-Flash 1017 named
According to reports, the models that benefited from distilling GPT-6 Astra were Kimi K3.1 Pro, released last week by Moonshot AI, and the upgraded V4.1-Flash 1017, launched by DeepSeek on October 17.
Chain News previously reported that Moonshot AI released Kimi K3.1 Pro on October 23, claiming it matched or came close to the strongest closed-source models on coding and agent tasks. Its officially reported Terminal-Bench 4.0 score was 65.7%, higher than GPT-6 Astra's 58.2%. The full model weights are expected to be made public by November 4.
Last month, OpenAI pointed to Moonshot AI in a wave of distillation activity
This is not the first time OpenAI and Moonshot AI have been linked to distillation. On September 30, OpenAI published a report revealing an adversarial distillation campaign that began in July, and attributed its core group to individuals associated with Moonshot AI.
OpenAI said the activity surged over July 24 and 25, when more than 4,000 users made 16,000 requests matching extraction patterns. It was fully blocked on July 28. OpenAI also said it could not determine whether all participants came from the same entity.
In September, the National Security Agency, the Federal Bureau of Investigation, and the Cybersecurity and Infrastructure Security Agency also issued a joint notice naming six companies—DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun, and Z.AI—for distilling U.S. models such as Claude, GPT, Gemini, and Grok on an industrial scale. Moonshot AI had previously denied using distillation to build Kimi K3.
Anthropic says safeguards have reduced distillation data extraction by more than 90%
Companies including Anthropic, OpenAI, and Google argue that distillation allows competitors to replicate advanced capabilities more quickly and cheaply.
On the 12th of this month, Anthropic said that after new anti-distillation measures went live, the amount of data distillers extracted from Claude fell by around 90% to 99%. However, Anthropic also admitted that it did not yet know how long the measures would remain effective, and expected distillation to be an ongoing game of cat and mouse.
This article is reprinted with permission from: (Chain News)
Original title: (DeepSeek and Moonshot AI reportedly used distillation of OpenAI models to train their latest models)
Original author: Elponcrab
The article “Online rumors claim DeepSeek and Moonshot AI used OpenAI extensively to train models, reigniting the distillation controversy” was first published by “Crypto City”