Mistral founder Arthur Mensch makes the case for open source AI. In a post on LinkedIn, he cautions companies against relying on closed AI models.
Mensch argues that companies selling closed models are storing more and more data, giving them a window into their customers’ business processes. Some AI labs “have a track record of tracking their most successful customers because of this information,” Mensch said.
He advises companies to store data in open systems, set their own access rules for AI, and build their own training models, even though “these efforts may seem difficult.” “Frontier AI can accelerate business growth, but if you don’t have it, you won’t grow,” Mensch wrote.
Mensch’s comments follow similar remarks from Palantir CEO Alex Karp, who urged companies to build their own AI models rather than relying on proprietary external solutions. Palantir also published a manifesto on secure AI in business. Among other things, it says, “To control your weight is to control your destiny. Your weight is a distilled form of hard-earned and accumulated institutional knowledge. When you let others control your weight, you are allowing them to shift the alpha part of your business into theirs.”
Mensch has a point, but he also has a business to run.
Mensch’s argument is valid, but requires context. Mistral is the only EU company with a relevant AI model and cannot compete with top-tier models such as GPT-5.6 Sol and Fable 5 in real-world performance. Mistral’s business model relies heavily on EU sovereignty. That’s because, even though about 30% of the company’s shares are held by U.S. investors, this is where the company stands to benefit the most. Large-scale general-purpose AI models also repeatedly outperform specialized models on specialized benchmarks, as long as relevant domain knowledge is part of the training data. Mensch discusses his book here.
However, recently published experiments on financial document analysis partially support him. Internal expertise not included in the training data for large models can provide an advantage.
Hedge fund Bridgewater and Thinking Machines Lab, a startup founded by former OpenAI CTO Mira Murati, used proprietary investor evaluations to fine-tune the open-source model Qwen3-235B. According to their own evaluation, for financial statements, the accuracy of the fine-tuned model reached 84.7 percent, while the best frontier model reached 78.2 percent. Operating costs were reduced by nearly 14 times.
This is not an independent comparison; both companies have a vested interest in selling their products. That too is just a snapshot. Companies like Anthropic and OpenAI could be back on top by simply purchasing or generating that kind of data for future training.
AI News Without the Hype – Curated by Humans
Subscribe to THE DECODER for ad-free reading, a weekly AI newsletter, exclusive six-a-year “AI Radar” frontier reports, full archive access, and access to the comments section.
Subscribe now
