France’s Mistral unveils biggest model in sovereign AI push

French startup Mistral announced Tuesday its latest open-weight model ML4, claiming enhanced performance and security capabilities for clients hoping to keep their data private instead of relying on closed models from the likes of Anthropic and OpenAI.

The new model, dubbed Chonk, lets users download and customise the parameters that determine how it computes.

It is Mistral’s first major product launch since it raised three billion euros ($3.4 billion) last month in Europe’s biggest technology fundraising ever.

The company is betting that companies and governments will increasingly want total control over their AI deployments, as opposed to using proprietary models that require handing over data to third parties and can be much more expensive.

“Enterprises across financial services, manufacturing and the public sector can govern the intelligence, data, compute and operations underpinning their critical infrastructure … without exposing their most valuable knowledge to anyone outside their own walls,” the company said in a statement.

Chonk is now available as a public preview for developers and government authorities for testing, before the model’s parameters, or “weights”, are released on October 27.

Mistral, considered Europe’s leading AI developer, was valued at over 21 billion euros in last month’s funding round — far below the valuations for US giants like OpenAI or the AI operations of Google, Meta and SpaceX.

Anthropic is reportedly preparing an initial public share offering in the coming weeks that could see it valued at $2 trillion.

But their so-called closed models that currently dominate the market are facing growing competition from open-weight models — largely from Chinese developers — that can be fine-tuned by clients for specific needs, and used at a fraction of the cost.

Mistral said Chonk’s scores on performance benchmarks so far make it one of the best open-weight models available, and “by far the best in Europe and the US”.

It said the model, which is still being refined, was built on 4,000 Nvidia chips over two months at its own data centres in Europe.

Customers can either download the model to their own servers, or access it via Mistral’s infrastructure, including in “our European sovereign region where data stays under EU jurisdiction to ensure data remains under their control”.