Report Says DeepSeek and Moonshot Used GPT-6 Astra Outputs to Train New AI Models
Two prominent Chinese artificial-intelligence developers are at the centre of a new dispute over how advanced AI systems are trained. A report by The Information, relayed by Reuters in coverage published on October 29, said DeepSeek and Moonshot AI relied heavily on outputs from OpenAI’s GPT-6 Astra model while developing recent releases. The account cited people described as having direct knowledge. It is a reported claim, not a publicly established finding that either company broke a law or contract. OpenAI, DeepSeek and Moonshot had not immediately replied to Reuters requests for comment at the time of the report.
The report identified two models: Moonshot’s Kimi K3.1 Pro, released in the preceding week, and DeepSeek’s V4.1-Flash 1017, launched on October 17. The allegation is that these models benefited from a technique known as distillation, in which the responses of one AI system are used as training material for another. That can be an ordinary engineering method when a developer uses its own systems or has permission to use the data. Controversy arises when a competitor’s outputs are allegedly collected or used in ways that conflict with access rules or commercial expectations.
What the reporting establishes—and what it does not
The public account names the developers, the alleged source model and the two recent releases. It does not provide a complete dataset, the volume of outputs involved, the accounts used to obtain them or a technical audit that readers can independently inspect. It also does not report a court judgment finding misconduct. Those gaps are not reasons to dismiss the allegation, but they are reasons to preserve its reported status. A headline stating that the companies definitely copied a model would go beyond what the available evidence establishes. A headline saying that a report alleges distillation is both more accurate and more useful.
Reuters also noted that OpenAI had previously said it linked part of a distillation campaign that began in July to people associated with Moonshot AI. That earlier company statement is relevant background, but it is not identical to the new report about particular model releases. Association with a campaign, a technical claim about training and a finding of a contractual violation are separate propositions. Readers should not merge them into a single certainty. The new report would become materially stronger if companies published verifiable details, acknowledged the activity or produced evidence contradicting it.
Why model distillation is controversial
Training a frontier AI model requires large investments in computing, data preparation, research and testing. Distillation can transfer some useful behaviour from an existing system into a smaller or different one by learning from its answers. Developers use the technique to make models cheaper to run or better suited to particular tasks. When the teacher model belongs to a rival, the economic argument changes: one company may have paid to create capabilities that another is accused of reproducing at lower cost. Whether any particular use is permitted depends on the facts, the terms of access and applicable law.
There are also technical limits to what can be inferred from similar outputs. Two AI systems may answer a question in comparable ways because they were trained on overlapping public material, because they use similar methods or because one learned from the other. Similarity alone is not conclusive proof of a specific training pipeline. Conversely, a model may have benefited from another system’s outputs without producing obvious copies in everyday use. Establishing what happened would require information about collection methods, training data, model-development records and relevant agreements that has not been made public in the reports reviewed.
What it means for users and developers
For people choosing an AI service, the immediate effect is uncertainty rather than a confirmed product shutdown or change in availability. The report does not say that Kimi or DeepSeek users will lose access to their tools, nor does it provide a new independent ranking of model quality. Businesses that rely on external AI systems may nevertheless watch the dispute because legal or commercial restrictions can affect vendor relationships. Developers evaluating these models should continue to test their own use cases and monitor official notices instead of treating an allegation as proof that a product is unsafe or unusable.
The story also reflects the increasingly competitive relationship between US and Chinese AI laboratories. Advanced models can be used to write software, analyse documents and support research, creating strong incentives to improve performance quickly. Leading companies have invested in technical measures to detect or restrict automated harvesting of model outputs. Those protections are part of a wider struggle over access, competition and the boundaries of legitimate learning from existing technology. The new allegation gives that debate a concrete set of model names and release dates, even though many operational details remain unknown.
The next significant development would be an on-record response from one of the companies, a published technical investigation or an identifiable legal action. Until such evidence appears, the current news should be described as a sourced report about possible distillation, not a settled finding of wrongdoing. The episode is likely to draw attention because it touches a central question for the AI industry: how much of a model’s competitive advantage can be reproduced by learning from its outputs, and under what conditions should that be allowed?
