Mistral releases Large 4, a 1-trillion-parameter multimodal model, with open weights due in three weeks

Mistral AI released Mistral Large 4 (ML4) on Tuesday, a multimodal model with one trillion parameters. It is not yet open-weight. For now, access is limited to a public guardrail endpoint, and the French lab says it will release the weights in three weeks, once safety testing is complete.

The model is nicknamed "Le Chonk" internally, a nod to its size. Mistral is positioning it as an alternative to both closed American models and open Chinese ones, a stance that echoes French president Emmanuel Macron's description of "a third way in AI."

What is confirmed

• Parameters: 1 trillion
• Modality: multimodal
• Access: public guardrail endpoint only; weights are not yet available
• Planned open-weight release: about three weeks out, after safety testing
• Context window: not yet disclosed
• Pricing: not yet disclosed
• Benchmarks: results still pending
• Training data cutoff: not yet disclosed

Training compute

According to Mistral VP Science Pierre Stock, ML4 was trained entirely on Mistral's own compute using only 4,000 NVIDIA GPUs. Stock claims this is "two to three times less than our Chinese competitors, and significantly less than the closed source competitors." These figures come from the company and have not been independently verified.

Performance claims

Mistral has published no benchmark results. The company says it hopes ML4 will be the best open-weight model, especially outside China. Stock also said focused training could let it beat closed models in specific areas where multimodal capability adds value. Those areas are cybersecurity, finance and chip design. Chip design matters to two of Mistral's main backers: ASML, which led its Series C, and Samsung, which led its Series D last month at a €21 billion valuation (about $24.39 billion).

Staged weight release

Mistral is holding back the weights while it works "with trusted partners and governments" so the open weights "can be used to defend, but not to [perform] malicious attacks," Stock told TechCrunch. Security concerns have been rising among Mistral's core enterprise and institutional customers. Stock also argued that open weights are easier to audit.

What this means

ML4 is Mistral's answer to doubts about its direction. The company recently began hosting Chinese models, and it has tried to show that this does not make it merely an inference provider. A trillion-parameter model trained in-house supports its case as a frontier lab.

The claims remain unproven. Without benchmarks, pricing or a context window, ML4 cannot yet be compared with closed frontier systems or with open-weight models from Chinese labs. The 4,000-GPU training claim, if it holds, would point to a notable efficiency advantage, but Mistral has not shared enough detail to assess it.

The three-week weight release is the next test. Publishing the weights would let outside researchers audit the model and run their own evaluations, and it would show whether the staged security process delivers on Mistral's pitch to risk-conscious enterprises and governments.