BEIJING / SAN FRANCISCO, 29 October 2026 — Chinese artificial intelligence developers DeepSeek and Moonshot AI relied heavily on outputs from OpenAI's powerful GPT-6 Astra model to train their latest models, according to a report published on Wednesday, reigniting one of the sharpest disputes in the global AI race: how far a rival may go in learning from a competitor's model.
The Information reported, citing people with direct knowledge of the matter, that the practice — known as distillation — helped shape Moonshot AI's Kimi K3.1 Pro model, released last week, and DeepSeek's upgraded V4.1-Flash 1017 model, launched on 17 October. Reuters reported the findings on 29 October 2026. OpenAI, DeepSeek and Moonshot AI did not immediately respond to Reuters' requests for comment, according to the report.
Distillation, in plain terms, means using the outputs of a larger, more capable model to train a smaller or newer model. Leading AI companies argue that the technique lets rivals reproduce advanced capabilities more quickly and at lower cost than building them from scratch — a claim at the heart of intensifying technology competition between the United States and China.
Illustration: Two stylized AI language model interfaces exchanging streams of text data.
What Happened
According to The Information, as reported by Reuters, Moonshot AI's Kimi K3.1 Pro and DeepSeek's V4.1-Flash 1017 both benefited from distilling GPT-6 Astra, OpenAI's flagship model. The report did not, in the Reuters account, publish internal documents or a detailed technical breakdown of how much of either model came from Astra outputs, and the companies involved did not confirm the account when Reuters sought comment.
That careful framing matters. Distillation is not a single, uniform act: it can range from studying a model's public answers to systematically harvesting large volumes of outputs to teach another system. What the report alleges is the heavier end of that spectrum — that the Chinese developers "relied heavily" on Astra outputs — rather than casual, incidental use.
The timing adds to the sensitivity. The report landed in the same week that OpenAI updated its public Model Spec, the document that sets out how it wants its models to behave, and shortly after rival Anthropic said new safeguards had sharply cut the data harvested in so-called distillation attacks on its Claude models. Together, the developments show an industry trying, in public and behind the scenes, to draw clearer lines around how models may learn from one another.
Illustration: A modern data center corridor with glowing server racks.
Background
The dispute is not new. In a blog post on 30 September, OpenAI said it had linked a core part of a distillation campaign that began in July to individuals associated with Moonshot AI, Reuters reported. That earlier disclosure set the scene for the latest report: OpenAI had already publicly connected Moonshot-linked activity to large-scale distillation, even before The Information's account of the Kimi and DeepSeek models.
Firms including Anthropic, OpenAI and Google have argued for some time that distillation allows rivals to reproduce advanced capabilities more quickly and cheaply, Reuters noted. Their concern is economic as well as technical. Training a frontier model can cost hundreds of millions of dollars in computing power, data and engineering time. If a rival can extract much of that capability by querying the finished model, the argument goes, the investment that created it is undermined.
Earlier in October, Anthropic said new measures designed to prevent rivals from copying the capabilities of its Claude AI models had reduced data harvested in distillation attacks by about 90% to 99%. That claim, reported by Reuters, illustrates the other side of the contest: model makers are now building defences — rate limits, detection systems and other safeguards — specifically to make large-scale harvesting harder.
DeepSeek and Moonshot AI sit at the centre of China's open-model ecosystem. Both have built reputations for releasing capable models at lower cost than many Western rivals, and both have attracted global attention from developers who value open weights and competitive performance. That is precisely why the distillation allegation carries weight beyond a single product launch: if the capabilities of leading Chinese models rest, in part, on outputs from a leading American model, it complicates claims about independent progress — and sharpens questions about what counts as fair competition.
Why It Matters
First, the report touches the competitive balance of the AI race. The United States and China are competing not only on model quality but on chips, data centres, talent and export controls. A finding that top Chinese models leaned on an American frontier model would feed arguments in Washington for tighter controls on model access, and in Beijing for faster development of home-grown alternatives.
Second, it tests the rules of the AI industry itself. Most model providers prohibit using their outputs to train competing models, and many treat large-scale distillation as a breach of their terms. But enforcement is difficult: outputs are, by design, meant to be read and used, and drawing a line between learning from a model and copying it is technically and legally murky. The latest report is likely to push that debate further into public view, and possibly into courts and regulators' offices.
Third, it matters for users and developers. If model makers respond by tightening access, adding more aggressive detection or restricting certain uses, the open, experimental character of AI development could narrow. If they do not, the largest investors in frontier models may question whether their spending can be protected at all. Either path shapes what tools developers can build on, and at what price.
Illustration: A human hand guiding a glowing AI robot head.
What Happens Next
Several questions remain open, and the answers will determine how far this story travels:
- Will OpenAI, DeepSeek or Moonshot AI respond in detail? None had immediately responded to Reuters at the time of its report. A denial, a confirmation, or a technical rebuttal would change the picture significantly.
- Will OpenAI or others take enforcement action? Model providers have previously used account blocks, technical safeguards and public disclosures rather than litigation, but the scale alleged here could prompt a stronger response.
- Will Anthropic-style defences spread? If Anthropic's reported 90% to 99% reduction in harvested data holds up, other labs may adopt similar measures, making large-scale distillation materially harder.
- Will regulators get involved? Distillation sits at the intersection of trade, intellectual property and national security policy, and governments on both sides have already shown interest in how AI capabilities spread.
- What do the next model releases show? Future releases from DeepSeek and Moonshot AI — and how they perform, and what their makers say about their training — will be watched closely for signs of a change in approach.
Separately, OpenAI's updated Model Spec, published on Thursday 29 October, reorganised its guidance around five values: faithfulness to human intent, harm minimisation, honesty and transparency, human agency, and care. OpenAI said faithfulness to human intent is the foundation of the framework, requiring models to understand what users, developers and OpenAI intend and value, rather than simply following instructions literally. The company said current models do not yet fully reflect the updated spec and that it is exploring new training methods to help models internalise the principles — a reminder that, even as labs police how others learn from their models, they are still refining how their own models should behave.
For now, the distillation report remains a report, not an adjudicated finding. It is attributed to The Information, citing people with direct knowledge, and reported by Reuters — and it should be read in that light. But its timing, its named models and its named source model make it one of the most consequential AI-governance stories of the week, and one that is unlikely to fade with a single news cycle.
Sponsored: explore today's featured sponsored offers
FAQ
What is AI model distillation?
Distillation is a training technique in which the outputs of one, usually larger and more capable, AI model are used to train another model. It can help a smaller model learn faster, but leading labs argue that large-scale distillation of a rival's model can reproduce its capabilities without the original investment.
Which models are named in the report?
Moonshot AI's Kimi K3.1 Pro, released last week, and DeepSeek's upgraded V4.1-Flash 1017 model, launched on 17 October 2026, are both said to have benefited from distilling OpenAI's GPT-6 Astra, according to The Information as reported by Reuters.
Have the companies confirmed the report?
No. Reuters reported that OpenAI, DeepSeek and Moonshot AI did not immediately respond to its requests for comment.
Why do AI companies object to distillation by rivals?
They argue it allows competitors to reproduce advanced capabilities more quickly and cheaply, undermining the investment in computing, data and research that went into the original model. Anthropic, OpenAI and Google have all made versions of this argument, according to Reuters.
What has OpenAI said before about Moonshot-linked distillation?
In a 30 September blog post, OpenAI said it had linked a core part of a distillation campaign that began in July to individuals associated with Moonshot AI, Reuters reported.
Sources
- Reuters, 29 October 2026: "DeepSeek, Moonshot used OpenAI's GPT-6 Astra to train latest models, The Information reports" — reporting The Information's account, the Kimi K3.1 Pro and DeepSeek V4.1-Flash 1017 models, the 30 September OpenAI blog post, and Anthropic's anti-distillation measures.
- Reuters, 29 October 2026: "OpenAI says human intent at center of updated AI model guidelines" — the updated Model Spec, its five values, and OpenAI's statement that current models do not yet fully reflect it.
- The Information, 29 October 2026, as reported by Reuters — the original report, citing people with direct knowledge of the matter.
Comments
Post a Comment