Mistral Unveils Large 4, Plans Open-Weight Release After Safety Testing

TL;DR

• Mistral AI released Mistral Large 4 on Tuesday, positioning the multimodal model as a European alternative to American and Chinese competitors.

• Nicknamed “Le Chonk,” the model has 1 trillion parameters and is initially available through a public endpoint with guardrails.

• Mistral plans to release the model’s weights in three weeks, following safety testing.

French AI company Mistral AI has introduced Mistral Large 4, a trillion-parameter multimodal model intended to strengthen its position in the race against leading American and Chinese AI developers.

Released Tuesday, the model reflects Mistral’s ambition to offer a European alternative in a market increasingly divided between proprietary systems and open-weight competitors. French President Emmanuel Macron has described that ambition as a “third way in AI.”

However, Mistral Large 4, or ML4, is not yet available as an open-weight model. Access is initially limited to a public endpoint with guardrails while the company completes safety testing.

Open-Weight Release Planned in Three Weeks

Mistral plans to make ML4’s weights available in three weeks, allowing users broader access to the model after the initial evaluation period.

Pierre Stock, Mistral’s vice president of science, told TechCrunch that the company will work with trusted partners and governments ahead of the release. The aim is to assess how the model can support defensive applications while addressing the potential for malicious use.

That process is particularly relevant to Mistral’s enterprise and institutional customers, whose concerns about AI security have grown in recent months.

Stock also argued that open-weight models offer an advantage for scrutiny because they are easier to audit. The planned release therefore sits at the intersection of Mistral’s openness strategy and its efforts to address security concerns.

Mistral Says ML4 Was Trained With 4,000 Nvidia GPUs

According to Stock, ML4 was trained entirely on Mistral’s own computing infrastructure using 4,000 Nvidia GPUs.

He said that figure was two to three times lower than the GPU resources used by Chinese competitors and substantially below those of closed-model developers.

The claim highlights Mistral’s emphasis on training efficiency as it competes with larger AI laboratories. However, benchmark results are still pending, leaving its relative performance unconfirmed.

Mistral hopes ML4 will rank among the strongest open-weight models once its weights become available. The company also expects focused training to help it compete with proprietary models in specific applications important to its customers.

Cybersecurity, Finance, and Chip Design Are Key Targets

ML4’s targeted use cases include cybersecurity, finance, and chip design, according to Stock. Mistral sees its multimodal capabilities as particularly useful in these specialized areas.

Chip design also connects the model’s development to two of the company’s major investors.

Dutch semiconductor equipment supplier ASML led Mistral’s Series C funding round, while Samsung led its Series D last month. The latter round valued Mistral at €21 billion, approximately $24.39 billion.

These relationships place semiconductor applications alongside enterprise security and financial services among the model’s strategic priorities.

Launch Reinforces Mistral’s Frontier AI Ambitions

The release follows questions about whether Mistral’s decision to host Chinese models signaled a shift toward primarily providing inference services.

Mistral has maintained that hosting other developers’ models does not replace its commitment to building advanced systems of its own. ML4 gives the company a new flagship model through which to demonstrate that ambition.

Its competitive standing will become clearer when benchmark results are published. For now, the launch establishes Mistral’s intended direction, with the planned weight release dependent on completion of safety testing.