PARIS — A Paris artificial intelligence laboratory on Friday released a large language model that it says performs on par with the leading proprietary systems from American companies, and made its full weights available for anyone to download and use commercially, a move that was greeted with enthusiasm by European officials and alarm by some safety researchers.
The model, called Clarté 3, has about 400 billion parameters and was trained over four months on a cluster of roughly 24,000 processors at a facility in the Paris region, Lumière Labs said. On a battery of standard benchmarks covering mathematics, coding, multilingual understanding and reasoning, the company reported scores within one or two points of the best closed systems, and ahead of them on tests in French, German, Italian and Polish.
“A continent of 450 million people cannot run its hospitals, its courts and its schools on models whose rules are written in California and can change overnight,” said the company’s chief executive, Dr. Antoine Rousseau, at a launch event attended by France’s digital minister. “Open weights are not a business model. They are a sovereignty model.”
Lumière, founded in 2023 by researchers from French universities and a large American technology company, has raised about 2.1 billion euros, roughly $2.3 billion, and is valued at around 14 billion euros. It earns revenue by selling hosted versions of its models and consulting services to companies and governments; releasing the weights for free, it argues, builds the ecosystem that buys those services.
The release comes as the European Union’s A.I. Act begins to apply fully to general-purpose models, requiring developers to document training data, assess systemic risks and report serious incidents. Lumière published a 140-page technical report alongside the model, including a description of the data sources, which it said excluded copyrighted books and paywalled news content, and an account of the safety testing it had carried out.
That testing did not satisfy everyone. “Once the weights are public, every safeguard the company built can be removed by anyone with a laptop and an afternoon,” said Dr. Elena Vasquez Moreno, a researcher at the European Centre for Algorithmic Transparency in Seville, who has argued for staged releases of frontier models. “The company is being transparent about a decision it cannot take back.” Lumière said its testing had found the model offered no meaningful uplift in dangerous capabilities over systems already freely available.
Industry reaction was swift. Within hours, the model had been downloaded more than 90,000 times, and several European cloud providers announced they would offer it as a hosted service. A German insurance group said it would begin evaluating Clarté 3 to replace an American model in its customer service systems. “The performance is one thing. The fact that it will still exist next year, unchanged, with the same license, is what matters to us,” said its chief technology officer, Ulrike Brandner.
Dr. Rousseau was asked how the company would compete with American rivals that spend many times its budget. “They have more chips,” he said. “We have more languages, more lawyers and, apparently, more patience.”
A version of this article appears in print on Oct. 11, 2026, Section A, Page 6 of the European edition with the headline: Paris A.I. Lab Releases a Freely Downloadable Model It Says Matches the Best Commercial Systems.
