Mistral’s New 1T Model Seeks to Surpass Open and Closed Competitors

Meta description: Mistral Large 4, a one trillion parameter multimodal model, enters the open vs closed AI race with 4,000 GPUs and open weights planned in three weeks.

Tags: Mistral AI, Large 4, open models, AI governance, European AI

Featured image: abstract technology illustration of a neural network lattice, no real people, no company logos, no text.

On Tuesday, Mistral AI released Mistral Large 4, a new large multimodal model that the French AI lab presents as a way to keep Europe in the mix while the race between open and closed AI accelerates. The release matters because it ties one trillion parameters, a public guardrail endpoint, and a promised three-week open-weights plan to a single launch, while benchmark results remain pending.

Background: A European Third Way in AI

The model arrives amid a growing divide between closed models that can be unplugged and open models that are often made in China. Mistral is positioning Mistral Large 4, abbreviated ML4, as an alternative to both sides. It follows what French president Macron has described as “a third way in AI.”

Mistral’s core audience includes enterprises and institutions, and security concerns have been mounting in recent months among those users. An open-weight model, whose parameters can be inspected and run by others, is easier to audit, said Pierre Stock, Mistral VP Science.

Development: From Guardrail Access to Open Weights

ML4 is not an open-weight model yet. For the time being, it can only be accessed via a public guardrail endpoint, a controlled hosted interface with safety checks. Mistral plans to make its weights available in just three weeks, after safety testing is complete.

The company frames the wait as a governance step. “In the meantime, we’ll work with trusted partners and governments to make sure that the open-source weights can be used to defend, but not to [perform] malicious attacks,” Stock told TechCrunch.

Nicknamed Le Chonk in reference to its one trillion parameters, ML4 is definitely not small. It is multimodal, meaning it is designed to work with more than one type of input or task, and Mistral says it aims to leapfrog American and Chinese rivals.

Latest: Training Footprint, Backers, and Frontier Lab Claim

Behind the model, Mistral says ML4 was trained entirely on Mistral’s compute, using only 4,000 NVIDIA GPUs. “which is two to three times less than our Chinese competitors, and significantly less than the closed source competitors,” Stock said.

Benchmark results are still pending. Mistral hopes ML4 will be best in class among open-weight models, especially outside China, but not only. It also says focused training could let the model outperform closed models in specific areas that are key to customers, especially where multimodal capabilities add value. The optimized use cases include cybersecurity and finance, and also chip design.

Those areas connect directly to Mistral’s backers. ASML led the company’s Series C, while Samsung led its Series D last month at a €21 billion valuation, about $24.39 billion. Chip design sits close to the AI hardware stack that supports model training and deployment.

Mistral has also been responding to questions about its identity. At the time, the company tried to convey that its decision to host Chinese models wasn’t a pivot into mere inference provider. With Le Chonk, Mistral believes it should still be considered a frontier lab as well.

Mistral Large 4 is now in a transitional stage: available through a public guardrail endpoint, not open-weight yet, with weights promised after three weeks and safety testing, and with performance claims waiting on benchmarks.