Mistral Launches ML4, Weight Specifications to Be Announced in Three Weeks

Le brief IA que les pros lisent chaque soir
Les 7 actus IA du jour, décryptées en 5 min. Gratuit.
Inclus dès l'inscription : notre sélection des meilleurs guides & comparatifs IA.
Choisis ton rythme
Gratuit · Pas de spam · Désabonnement en 1 clic
Mistral AI presents ML4, a trillion-parameter multimodal model initially offered via a public access point with safeguards. The weights are expected to be released after security testing, while the company emphasizes its sector priorities and proprietary training methods.
Mistral claims cutting-edge lab status after its Series C and D rounds
Samsung led Mistral's Series D last month, valuing the company at €21 billion, or approximately $24.39 billion. Chip design, cited among the intended uses of ML4, is crucial for ASML, which led the Series C. Mistral clarified that hosting Chinese models does not signify a shift towards being merely an inference provider. With ML4 Le Chonk, the company believes it should be regarded as a cutting-edge technology lab. Benchmark results for the new model are still pending.
Access via API, weight release announced, and focus on security
ML4 is not yet available for access to its weights and is currently only available through a public endpoint equipped with protective measures. Mistral plans to release the weights in three weeks, once security checks are completed. Pierre Stock states that the company will work with trusted partners and governments to ensure that open weights are used for defense rather than malicious attacks. Security concerns have increased in recent months among targeted companies and institutions. According to him, an open-weight model is easier to audit.
Announced capabilities, training on 4,000 GPUs, and business targets
Mistral AI has launched ML4, a large multimodal model nicknamed Le Chonk in reference to its trillion parameters. The entire training of the model was conducted on Mistral's internal infrastructure, utilizing 4,000 Nvidia GPUs, a number that Pierre Stock estimates is two to three times less than that used by Chinese competitors and significantly lower than that of closed-source players. The company aims to gain an edge over its American and Chinese competitors and aspires for ML4 to position itself as the best open-weight model in its category, particularly outside of China. With specialized training, ML4 could surpass certain closed models in specific sectors deemed strategic for clients, such as cybersecurity, finance, or chip design. This strategy is set against a backdrop described as increasingly polarized between closed and open models, while Mistral promotes ML4 as an alternative to both, in the spirit of the "third way" mentioned by Emmanuel Macron.
Brief IA — L'actualité IA en français
L'essentiel de l'actualité de l'intelligence artificielle, décrypté et expliqué chaque jour.