Chinese artificial intelligence developers DeepSeek and Moonshot AI relied heavily on outputs from OpenAI’s GPT-6 Astra model to train their latest systems, according to a report by The Information, intensifying concerns among leading AI companies over the use of model distillation to replicate advanced capabilities at a lower cost.
The report, published on Wednesday and citing people with direct knowledge of the matter, said Moonshot AI’s Kimi K3.1 Pro, released last week, and DeepSeek’s upgraded V4.1-Flash 1017 model, launched on October 17, benefited from distillation using OpenAI’s model.
The reported use of Astra outputs has not been independently established in the information provided by the companies.
Register for the next Tekedia Mini-MBA.
Register for Tekedia AI in Business Masterclass.
Join Tekedia Capital Syndicate and co-invest in great global startups.
Distillation involves using the outputs of a more capable AI model to train another model to reproduce similar responses or behaviors. Instead of developing every capability through an entirely independent training process, developers can use responses generated by a leading system to help a smaller or newer model learn how to perform particular tasks.
The technique has become a source of tension in the AI industry because it can reduce the cost and time required to develop competitive models. Companies that invest heavily in training advanced systems argue that competitors may be able to benefit from their work without bearing comparable research, computing and development expenses.
The latest report adds to the growing dispute over where legitimate model development ends and the unauthorized extraction of a rival’s capabilities begins.
The allegations emerge as Chinese AI developers seek to compete with leading American model providers in coding, reasoning and other advanced applications. Distillation can help developers improve model performance more efficiently, particularly when they can access large volumes of high-quality outputs from a more capable system.
For companies such as OpenAI, Anthropic and Google, however, the practice raises concerns about the commercial value of their research investments. If competitors can reproduce important capabilities by systematically querying existing models, the financial advantage of developing a frontier system could be reduced.
The US government said in a September that there is a great distillation campaign by Chinese companies against American AI companies. That statement is separate from The Information’s latest reporting about the use of GPT-6 Astra to train the Chinese company’s newer model.
The existence of a reported distillation campaign does not reveal the full extent of any model’s reliance on another system, whether specific terms of service were violated or whether all the resulting capabilities were obtained through that process.
The broader concern for AI developers is that the economics of model development could change if access to a competitor’s outputs becomes a substitute for some of the costly work required to train and refine a frontier model.
The technique also does not necessarily allow a rival to reproduce an entire model. Outputs can provide useful training signals, but they do not automatically transfer the original system’s architecture, training data, internal parameters, or every capability. The effectiveness of distillation is determined by the quality and diversity of the examples and the training process used to incorporate them.
Nevertheless, large-scale extraction could give competitors a faster route to particular capabilities, potentially increasing competitive pressure on companies that spend billions of dollars developing advanced models.
AI Companies Strengthen Safeguards Against Model Extraction
OpenAI is not alone in trying to limit the practice. Anthropic and Google have also noted that distillation can allow competitors to reproduce advanced AI capabilities more quickly and cheaply.
Earlier this month, Anthropic said new measures designed to prevent rivals from copying the capabilities of its Claude models had reduced data harvested through so-called distillation attacks by approximately 90% to 99%.
The reported reduction suggests that defensive measures can make large-scale extraction more difficult, although the effectiveness of such safeguards depends on the methods used and the ability of attackers to adapt.
Model providers can monitor unusual query patterns, restrict access to certain capabilities, impose usage limits, and investigate activity that appears designed to collect training data at scale. These measures can raise the cost of extracting a model’s outputs, but they also create operational challenges for legitimate developers and businesses that depend on access to AI systems.
The industry faces a difficult balance. Broad access to models supports experimentation, product development, and competition, while unrestricted access can make it easier for third parties to collect large datasets of outputs for purposes that providers may not have authorized.
Distillation itself is not inherently improper. AI developers routinely use techniques that transfer knowledge between models, and smaller models can be trained from outputs generated by larger ones. The controversy centers on the circumstances of the data collection, the permissions governing access, and whether a competitor has systematically extracted capabilities in breach of applicable restrictions.
Those distinctions are becoming increasingly significant as AI models become commercial products and their developers seek to protect the returns on expensive research and infrastructure investments.
Thus, the reported use of Astra outputs by DeepSeek and Moonshot AI raises questions about how effectively OpenAI can protect the capabilities of its most advanced systems while maintaining access for customers and developers.
The issue has now gone beyond a model’s ability to match the performance of a leading rival to how that performance was achieved, what resources were required, and whether the methods used comply with the conditions under which competing systems are accessed.