Two of China’s leading AI labs reportedly leaned heavily on outputs from OpenAI’s most powerful model to train their newest systems, putting a familiar dispute over model distillation back at the centre of the AI race. DeepSeek’s upgraded V4.1-Flash model and Moonshot AI’s Kimi K3.1 Pro, both released this month, are said to have benefited from distilling GPT-6 Astra. Neither lab nor OpenAI had responded to requests for comment when the report appeared, and the account remains unconfirmed by the companies involved.
What the report says
The account says people with direct knowledge described heavy reliance on Astra outputs in training the two new models. It names DeepSeek’s V4.1-Flash 1017, launched on 17 October, and Moonshot’s Kimi K3.1 Pro, released the previous week. The report does not say how much of either model’s training data came from Astra, nor publish documents supporting the claim.
That gap matters, because distillation covers a range of practices. At one end, a lab may use a rival’s answers as one training signal among many. At the other, it may harvest outputs at scale to reproduce a competitor’s capabilities. The report does not place either lab on that spectrum, so its claim should be treated as reported, not established.
What distillation means
Distillation means feeding a model prompts, collecting its responses, and using them to train a different model. The student can absorb much of the teacher’s capability at a fraction of the cost of training from scratch. Labs argue this lets rivals reproduce capabilities more quickly and cheaply, which is why it has become a flashpoint as United States-China competition intensifies.
The technique is also difficult to police. Model outputs are not the same as source code or weights, and labs disagree about where legitimate learning from public answers ends and copying begins. That disagreement is now a commercial question as well as a technical one, because the models at issue are being sold to businesses and pitched to investors.
The two models at issue
DeepSeek went first, launching V4.1-Flash 1017 on 17 October as an update focused on sharper task performance and better agentic coding. Agentic coding means the model plans and carries out multi-step programming work with less hand-holding, rather than simply suggesting code. DeepSeek kept its signature approach intact, with the model remaining open-weight and low-cost.
Moonshot followed with Kimi K3.1 Pro, a substantial upgrade over a predecessor released only a few months earlier. The model packs about 2.8 trillion parameters and scored 55 on the Artificial Analysis Intelligence Index, placing fourth overall. It also beat GPT-6 Astra on some agentic benchmarks, including Terminal-Bench 4.0, a result that will draw attention if the distillation claim holds.
A history of accusations
These two names have appeared in a distillation dispute before. In February 2026, Anthropic accused both Moonshot and DeepSeek of running distillation attacks on Claude, alleging the labs used fraudulent accounts to extract outputs. Anthropic cut off paying customers it linked to activity, a step it said was worth 5% to 7% of revenue, as reported in this site’s coverage.
OpenAI has also raised the issue. In a blog post on 30 September, it said it had linked a core part of a distillation campaign that began in July to individuals associated with Moonshot AI. Earlier this month, Anthropic said measures designed to prevent rivals copying Claude’s capabilities had reduced data harvested in distillation attacks by about 90% to 99%.
Why It Matters
The stakes are immediate for Moonshot in particular. The lab’s valuation reportedly reached about $50 billion by October 2026, ahead of a planned listing in Hong Kong, and a public offering invites detailed questions about training practices. Distillation allegations are the kind of issue prospective shareholders tend to notice, and they can follow a company well beyond its listing.
For DeepSeek, the low-cost, open-weight playbook has been its calling card since it first unsettled the market. If questions about where its training signal comes from stick, that cost advantage may draw more scrutiny than praise. The wider question is whether OpenAI follows Anthropic and makes a public case over Astra, moving the debate from industry argument towards formal policy.
Background
DeepSeek rose to prominence by offering capable open-weight models at a fraction of the cost of its American rivals, a formula that forced the whole industry to rethink pricing. That history explains why a claim about borrowed training signal cuts so close to the lab’s identity, and why its rivals guard their outputs so closely.
Source: Reuters report and further coverage
0 Comments