Mistral’s new 1T model aims to leapfrog closed and open rivals

French AI lab Mistral AI has unveiled Mistral Large 4 (ML4), a multimodal model with roughly 1 trillion parameters nicknamed "Le Chonk." The company is initially offering ML4 through a guarded public endpoint, and plans to release the model weights in about three weeks after completing safety tests.

By AI Newsroom· Reviewed by Pranav, Founder & Editor-in-ChiefPublished 2 minutes agoUpdated 2 minutes ago0 views

Mistral’s new 1T model aims to leapfrog closed and open rivals

Why It Matters

ML4 positions Mistral as a European contender between closed, U.S.-style models and many open models developed in China, aiming to balance accessibility with security. Its release underscores growing industry debate over open weights, auditability and responsible access to powerful AI systems.

Key Facts

Model name: Mistral Large 4 (ML4), nicknamed Le Chonk
Scale: Approximately 1 trillion parameters
Access at launch: Public guardrail endpoint (weights not yet open)
Planned weights release: About three weeks after safety testing
Training compute: Trained on 4,000 Nvidia GPUs

Mistral AI announced Mistral Large 4 (ML4), a new large multimodal model that it says contains about 1 trillion parameters and has been nicknamed "Le Chonk." The company positioned ML4 as an alternative to both fully closed models and many open-weight models emerging from China, echoing French president Emmanuel Macron's description of a "third way in AI." At launch, ML4 is accessible only through a public guardrail endpoint while Mistral completes safety testing.

Mistral plans to publish ML4's weights roughly three weeks after the safety review is finished, according to the company. Mistral VP of Science Pierre Stock said the firm will collaborate with trusted partners and governments to try to ensure the open weights can be used defensively rather than for malicious purposes. Stock also noted that open weights make auditing easier, a recurring concern for Mistral's enterprise and institutional customers.

On infrastructure, Mistral said ML4 was trained entirely on its own compute using 4,000 Nvidia GPUs, which the company claims is two to three times fewer than some Chinese competitors and materially less than some closed-source rivals. Benchmark results for ML4 are still pending; Mistral hopes the model will lead among open-weight models outside China and may outperform closed models in certain customer-focused tasks where multimodal abilities matter.

Mistral has highlighted specific use cases for ML4 that align with its customers' needs, naming cybersecurity, finance and chip design. The firm's emphasis on chip design reflects the interests of two major backers: ASML, which led Mistral's Series C, and Samsung, which led the Series D round that valued the company at €21 billion (about $24.39 billion). Mistral says it intends to retain its identity as a frontier AI lab rather than pivoting solely to inference services.