What Happened to Le Chat? Mistral’s Vibe Rebrand, Tested (2026) - MRKT3.0
Skip to content
European AI News & Startups, Startups
What Happened to Le Chat? Mistral’s Vibe Rebrand, Tested (2026)
Akos Szima
July 20, 2026
Takeaways
Mistral retired Le Chat, its best-known consumer brand, on 28 May 2026 and rebuilt it as Vibe, an agent that works across chat, work and code.<br>The company is reportedly raising at a roughly €20bn valuation, nearly double its September price, on annualised revenue of only about $400M.<br>When asked to recall financial data unprompted, Vibe invented a $337 share price for Cohere, a private company with no stock at all, which is the clearest case going for keeping a human in the loop.
Mistral has retired its most recognisable product at the height of its fame, and the replacement is better than the coverage suggests. On 28 May 2026, Le Chat became Vibe , a three-tab work agent (chat, work, and code) built for enterprises rather than casual users. The rebrand reads as a retreat from the consumer chatbot war with ChatGPT , but the product underneath is more capable than the name change implies, with one alarming exception.<br>The team at MRKT3.0 ran it through our tests to find where it holds up and where it makes things up.<br>Why Did Mistral Kill Its Best-Known Brand?<br>Le Chat was Mistral’s biggest consumer success. The mobile app hit one million downloads in its first 14 days, helped along by an endorsement from France’s President Emmanuel Macron . So why did Mistral retire Le Chat?<br>Mistral AI released Mistral Vibe 2.0, an upgraded SWE agent CLI with subagents, skills support and new unified agent modes. Now available on Team and Pro plans.
Terminal Testing Time 👀 https://t.co/eD1Rwekgj3 pic.twitter.com/PeNLLj28sP<br>— 🚨 AI News | TestingCatalog (@testingcatalog) January 27, 2026
Vibe’s pitch is an AI that does work rather than chats. A “Work Mode” connects to Google Workspace , Outlook , SharePoint , Slack and GitHub and runs multi-step tasks, from drafting reports to processing email, once the user approves. There is web search, a canvas editor, a code interpreter, document OCR and image generation. That last one runs on FLUX from Black Forest Labs , the German image lab, which is a little nod to Mistral’s mission of European Sovereign AI.<br>Mistral explained that the rebrand was focusing on enterprise work agents , which is where a significant proportion of revenue is in the AI industry. But, perhaps the less charitable reading is that the consumer chatbot war against ChatGPT could not be won, so Mistral is moving the fight to ground where distribution matters more than brand. Probably both are true.<br>Who Pays for Mistral, and What Do They Want?<br>Mistral has raised about €3.5bn in equity and debt since 2023. The September 2025 Series C brought in €1.7bn at an €11.7bn valuation and made ASML , the Dutch lithography monopolist, its largest shareholder with a €1.3bn cheque. No other frontier lab has a chip-equipment maker at the top of its cap table. Whether that is far-sighted industrial strategy or Europe’s corporate champions closing ranks around a national asset is a fair question, and nobody in Paris seems eager to answer it.<br>In June, Bloomberg reported that Mistral was in early talks to raise about €3bn at a valuation near €20bn , almost double the Series C figure. Terms were not final when this article went to press, and later reports put the figure higher still. Revenue has meanwhile grown from roughly $20M to more than $400M of annualised run-rate in a year , and the company says it wants $1bn by year-end .<br>In March, Mistral borrowed $830M , debt rather than equity, to buy 13,800 NVIDIA GB300 chips for a new data centre outside Paris. Banks lending against AI revenue says something about how predictable that revenue has become. It also says something about sovereignty: Europe’s champion runs on American silicon , bought with leverage, and its largest shareholder is the one company whose machines make that silicon possible.<br>How Good Is Mistral Large 3?<br>Mistral’s current flagship, Mistral Large 3 , is a 675-billion-parameter mixture-of-experts model with open weights under Apache 2.0, a permissive licence at a size where most rivals stay closed. On paper it is among the strongest open non-reasoning models available: about 85.5% on a multilingual MMLU, 93.6% on MATH-500, 92% on HumanEval, and a debut at second place among open non-reasoning models on the LMArena leaderboard.<br>On GPQA Diamond, a benchmark built to resist memorisation, Large 3 scores about 44%. Open "thinking" models like DeepSeek-V3.2 and Kimi K2 roughly double that. Output speed of about 38 tokens per second is slow for the class. In short, Large 3 is a very good multilingual workhorse and a mid-table deep reasoner.<br>What Did Vibe Say Happened to Le Chat?<br>The team at MRKT3.0 were curious about Mistral’s new model, so we ran a short test.<br>Our standard test set, the same five prompts we put to every model in this series,...