Mistral AI releases Mistral Large 4, a 1 trillion-parameter multimodal model with planned open-weight release

French AI lab Mistral AI released Mistral Large 4 (ML4), a large multimodal model with 1 trillion parameters, positioning it as an alternative to both American closed models and Chinese open models. Mistral described ML4 as part of a 'third way in AI' amid growing divides between closed and open approaches.

ML4 is not yet an open-weight model, but Mistral plans to release its weights publicly within three weeks, pending completion of safety testing. In the interim, access is provided via a public guardrail endpoint. Mistral’s Vice President of Science, Pierre Stock, emphasized the company’s commitment to enabling defensive use while preventing malicious applications of the open-source weights.

ML4 was trained entirely on Mistral’s compute infrastructure using 4,000 Nvidia GPUs, which Mistral claims is two to three times fewer than some Chinese competitors and significantly less than closed-source competitors. Benchmark results for ML4 have not yet been published.

Mistral highlighted optimized use cases for ML4, including cybersecurity, finance, and chip design. The company noted that multimodal capabilities could provide value in these domains, particularly for key customers such as Dutch semiconductor firm ASML and Samsung, which led Mistral’s Series C and Series D rounds, respectively. Mistral’s Series D was announced at a €21 billion valuation (approximately $24.39 billion).