Mistral accused of distilling DeepSeek model and falsifying benchmarks
A departing Mistral employee allegedly accused the company of distilling DeepSeek-v3 to create its Mistral Small 3.2 model while falsely presenting it as a reinforcement learning achievement and misrepresenting benchmark results. Analysis of word frequency patterns by an external observer suggested high output similarity between the models, consistent with distillation. Mistral has not publicly responded to these allegations.
What happened
A departing Mistral employee allegedly sent an email on Twitter accusing the company of distilling DeepSeek-v3 to create Mistral Small 3.2, then misrepresenting the result as a reinforcement learning success and falsifying benchmark test results.
Twitter user Sam Peach analysed Mistral Small 3.2 and DeepSeek-v3 outputs for unusual word frequency patterns (called 'Slop'), finding high similarity between the models that would be unlikely from independent training, suggesting distillation.
Reported impact
The allegations undermine Mistral's reputation as Europe's leading open-source AI company, particularly given the company's high profile in the market and recent $10 billion valuation.
If accurate, the claims suggest Mistral misled the public about its model development approach and capabilities.
Index analysis
Distillation itself is a standard technique used by many AI companies to improve model performance, but it requires transparent labelling according to industry observers cited in the reports.
The core allegation is that Mistral concealed the use of distillation from DeepSeek while claiming independent reinforcement learning success.
Governance controls
Mistral has not responded officially to the allegations as of the reported date.
Classification
Sources and evidence
This record was researched and written by the Index, based on the following original reporting.