DeepSeek and Moonshot AI trained models using OpenAI’s GPT-6 Astra via distillation

How distillation was used

According to a report by The Information, DeepSeek and Moonshot AI used OpenAI’s GPT-6 Astra model to train their latest systems, Kimi K3.1 Pro and DeepSeek V4.1-Flash, through a process called distillation. Distillation allows developers to replicate capabilities of a more advanced model by collecting its responses to large volumes of prompts and using them to train their own systems at a lower cost.

Moonshot AI released Kimi K3.1 Pro last week, while DeepSeek launched an updated version of its V4.1-Flash model on 17 October.

OpenAI’s accusations and industry concerns

OpenAI had previously accused Moonshot AI of running a coordinated campaign in July to extract hidden reasoning from its models. The company said it detected activity involving more than 15,000 users linked to the campaign, which it shut down by 28 July. OpenAI shared its findings with other AI labs, including Anthropic and Google.

Anthropic, OpenAI, and Google have argued that distillation enables competitors to reproduce advanced AI capabilities without equivalent investment in training. Anthropic claimed its new safeguards reduced data harvested from distillation attacks on its Claude models by 90% to 99%.

Benchmark claims and unverified results

Moonshot AI stated that its Kimi K3.1 Pro model outperformed GPT-6 Astra on a coding benchmark, while DeepSeek introduced V4.1-Flash as a lower-cost alternative to US-developed models. Neither company’s benchmark results have been independently verified.

Broader implications for AI development

The use of distillation has become a flashpoint in US-China technology competition, with US agencies identifying Chinese firms allegedly targeting American AI models. The practice highlights a challenge for leading AI labs: their publicly available models can inadvertently serve as training material for competitors seeking to close the gap.

What this means

Lazyfounder analysis — our interpretation, not reported fact.

This story underscores the risks and complexities of model distillation for AI developers. For startups and established labs, it signals that even advanced safeguards may not fully prevent competitors from leveraging their work. The lack of independently verified benchmarks from Moonshot and DeepSeek also raises questions about how to assess claims of model superiority, particularly in a competitive and geopolitically charged environment. Operators may need to factor in the possibility of distillation when designing access controls or APIs, especially if their models hold strategic value.

Related on Lazyfounder

• OpenAI
• Google
• More ai news