Chinese AI labs DeepSeek and Moonshot have reportedly used outputs from OpenAI's most powerful model, called Astra, to help train some of their recent releases, according to people with direct knowledge of the matter.
Reported allegation · OpenAI, DeepSeek, Moonshot
Did Chinese labs train on OpenAI's most powerful model?
Anonymous sources say DeepSeek and Moonshot used outputs from OpenAI's Astra to help train some recent releases. The details are limited, and neither company's response is included.
How distillation is alleged to work
Step 1
Astra outputs
Responses from OpenAI's most capable model.
→
Step 2
Distillation
Outputs are used to train a separate model, transferring capability without matching the original training cost.
→
Step 3
Student model
The resulting model, which the report links to DeepSeek and Moonshot releases.
What the report says
Astra has become a major target for Chinese AI developers.
Distillation can transfer capabilities at lower cost than full training.
OpenAI's usage policies generally bar using outputs to build competing models.
The allegation touches on whether frontier models can be replicated through public interfaces.
What remains open
The specific models trained and the extent of the use are not described.
How the outputs were obtained is unclear.
How much the outputs contributed to final models is unknown.
Whether OpenAI has taken action against the companies is not stated.
Key takeaway
If confirmed, the episode could push AI companies to tighten access to their most advanced systems.
Access controls
Pressure to limit how public interfaces expose frontier models.
Terms of service
Cross-border enforceability of restrictions on competing models is hard to test.
Policy attention
Possible interest from policymakers focused on AI competition and intellectual property.
Continue reading The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.
Already purchased? Sign in ✓ Signed in — this article isn’t included in your current plan.Unlocking the full article...