Mistral launches Large 4, a trillion-parameter multimodal AI model

French AI lab Mistral AI has released Mistral Large 4, a new multimodal model with one trillion parameters. The company has nicknamed the model “Le Chonk” and is positioning it as a European alternative to both closed US systems and open models often developed in China.

Mistral Large 4 is not yet available as an open-weight model. For now, access is limited to a public guardrail endpoint. Mistral plans to make the weights available in three weeks, once safety testing has been completed.

A staged route to open weights

Pierre Stock, Mistral’s VP of Science, said the interim period will be used with trusted partners and governments to help ensure that the weights can support defence rather than malicious attacks. The company’s approach responds to security concerns among enterprises and institutions, which are central to Mistral’s customer base.

Stock also argued that open weights can make a model easier to audit. That creates a tension Mistral is trying to manage: broader inspection and deployment possibilities on one hand, and safeguards against misuse on the other.

The launch follows Mistral’s sovereign AI infrastructure funding, which underscored Mistral’s focus on sovereign AI infrastructure and provides context for its effort to remain a frontier-model developer rather than only an inference provider.

Compute efficiency and target workloads

Mistral said it trained Large 4 entirely on its own compute using 4,000 NVIDIA GPUs. Stock described that total as two to three times lower than that of Chinese competitors and significantly lower than that of closed-source competitors.

Benchmark results are still pending, so the company’s performance claims remain to be tested. Mistral nevertheless expects Large 4 to rank among the strongest open-weight models, particularly outside China, and says focused training could help it outperform closed models in selected tasks.

Enterprise use cases shape the model’s pitch

The company identifies cybersecurity, finance and chip design as optimized use cases for the model, where multimodal capabilities may add value. Chip design is especially relevant to two of Mistral’s backers: ASML led its Series C, while Samsung led a Series D round last month at a €21 billion valuation, or about $24.39 billion.

For businesses, the practical implication is to distinguish the current guarded endpoint from the planned open-weight release, and to evaluate safety testing, audit needs and workload-specific performance before selecting Large 4 for sensitive deployments.