Notes from the AI Now Summit by Mistral

Koen van Gilst's personal reflections on attending Mistral's AI Now Summit in Paris. Published May 29, 2026. 4 min read. Also published on LinkedIn and Hacker News.

Key Points

1. Mistral's evolution beyond models — The company is building a full AI stack including compute (a 40MW data center in Paris, with more including one in Sweden), platforms, and consultancy, not just models.

2. Partnerships took center stage — The summit emphasized collaborations with ASML, BNP Paribas, and Amazon's Alexa+ rather than announcing new models or tech innovations, which the author found "disappointing."

3. Agentic systems require a "harness" — Pieter Stock argued that "the model alone isn't enough." A harness adds context, persistence, learning, and reasoning — the latter being essential for backtracking, error recovery, and transparency. Skills let organizations capture best practices developed cooperatively with AI agents.

4. Specialized small models as core strategy — Examples included Document AI for OCR (EU Patent Office), Voxtral for multilingual voice (Amazon's Alexa+ in Europe), and Robostral for industrial robotics (ASML). The author notes that "speed and efficiency are becoming as important as raw capability."

5. Sovereignty and on-prem deployment as differentiators — BNP Paribas runs Mistral models on-prem for KYC in Belgium, keeping sensitive data in-house. Abanca handles over 1 million customers through agent orchestration. The author frames this as "a good alternative to relying on US hyperscalers" for regulated European industries.

6. Humanities use case — A research team from the Austrian Academy of Sciences fine-tuned Codestral (a Mistral coding LLM) to read ancient papyrus fragments. This makes a collection of 180,000 documents accessible — a task estimated to take over 2,000 years without AI. The author calls it a "beautiful example of how AI can also help the humanities."

Notable Quotes

• "Mistral is no longer just a model company."
• "the model alone isn't enough"
• "speed and efficiency are becoming as important as raw capability"
• "a good alternative to relying on US hyperscalers"

Core Themes

• European AI sovereignty — Mistral positioned as a European full-stack alternative to US tech giants, with the author remarking the days of relying on US companies "is coming to an end."
• Enterprise pragmatism over AGI race — The author's takeaway: Mistral may aim "not to win the race for AGI" but to deliver real return on investment now through open models, on-prem deployment, and enterprise partnerships.
• Efficiency vs. raw capability — A recurring theme that smaller, specialized models can outperform general-purpose ones in specific, token-heavy applications.
• Data privacy and regulatory compliance — On-prem deployment as a selling point for regulated European industries handling sensitive data.
• AI for cultural heritage — The Codestral/papyrus project stood out as a non-commercial, humanities-focused application.

Author's Voice

Van Gilst writes in a reflective, first-person blog style — personal and candid. He expresses mild disappointment at the lack of new model announcements but emerges with a clearer understanding of Mistral's positioning. The tone is generally optimistic about European AI, and he closes with gratitude for the invitation, noting the venue near the Louvre and speakers on a fashion week catwalk.