Reports have surfaced suggesting that the French AI startup Mistral may have used a “distilled” version of Chinese-developed models, specifically DeepSeek, to train its own large language models. The claims, which emerged from a whistleblower, question Mistral’s public narrative of success through Reinforcement Learning (RL) and a unique Mixture of Experts (MoE) architecture.
The core of the controversy centers on “distillation,” a common technique in AI development where a smaller model is trained to mimic the behavior of a larger, more powerful “teacher” model. An online report by a whistleblower claims to have found “linguistic fingerprints” linking Mistral’s output to that of DeepSeek, which had previously demonstrated strong reasoning capabilities. This practice of distilling from another company’s model, particularly when the original model’s output is generated via API, blurs the lines of intellectual property and can lead to legal disputes, as seen in a separate case between OpenAI and DeepSeek.
While Mistral and DeepSeek have not issued public statements on the matter, the allegations bring the training practices of fast-rising AI firms into the spotlight. The use of synthetic data—often generated by other AI models—is a growing concern, as it can lead to the replication of biases and errors across generations of models, a phenomenon some researchers have likened to a “Xerox of a Xerox” effect.