Paris Soir - France's Mistral unveils latest model in sovereign AI push

The new model lets users download and customise the parameters, or "weights", that determine how it computes to their own servers.

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.

"That one took some groundwork," Mistral's chief executive Arthur Mensch posted on X.

"Trained and served on our own compute, and RL shows no sign of saturation," he wrote, referring to reinforcement learning to further develop AI models, and the possibility that adding further data starts to produce diminishing performance returns.

Open-weight models are also touted as more secure, a growing concern after several incidents of AI agents going rogue by hacking databases or otherwise doing things they were not supposed to do.

"The cyber defence capabilities will enable enterprises and governments to defend themselves against threat actors that are jailbreaking closed models to perform cyberattacks," Mistral's Chief Scientist Guillaume Lample said in the statement.

- Data 'under their control' -

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 including DeepSeek and Moonshot.

These AI versions can be fine-tuned by clients for specific needs on in-house infrastructure, and used at a fraction of the cost.

Mistral's new model also comes as a US open-weight rival, Reflection AI, announced its own widely anticipated Beam model on Monday, also promising cheaper operating costs.

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".

Mistral said the model was trained on more than 160 languages, and can answer prompts and questions in each one natively.

js/rl

X.Francois--PS