Mistral AI released Large 4 in October 2026. This flagship model uses a mixture-of-experts architecture with 1000 billion parameters, where 52 billion remain active. The company also launched Le Chonk in September 2026, which utilizes 49 billion active parameters from a 1 billion parameter total to provide high performance within a compact architecture. Engineers trained Le Chonk using 3800 NVIDIA Grace-Blackwell GPUs. This model remains the only significant open-weight model not originating in China. Earlier in 2026, Mistral released Medium 3.5 in April, a 128 billion parameter model under a modified MIT license. In March 2026, Mistral Small 4 arrived with 119 billion parameters and 128 experts, providing both text and image inputs via an Apache-2.0 license. The company also released Voxtral Realtime in February 2026, which contains a 4 billion parameter model under an Apache-2.0 license specifically designed for real-time speech transcription tasks. In December 2025, Devstral 2 released as a 123 billion parameter dense model with a modified MIT license. Devstral Small 2 arrived in December 2025 with 24 billion parameters under Apache-2.0. A 24 billion parameter reasoning model, Magistral Small, arrived in June 2025.
| Model Name | Release Date | Parameters (Total/Active) | License |
|---|---|---|---|
| Large 4 | October 2026 | 1000B / 52B | Mistral Ecosystem Open-Weight |
| Medium 3.5 | April 2026 | 128B | Modified MIT |
| Mistral Small 4 | March 2026 | 119B / 4 active per token | Apache-2.0 |
| Voxtral Realtime | February 2026 | 4B | Apache-2.0 |
| Devstral 2 | December 2025 | 123B | Modified MIT |
Development and coding agents
Mistral provides Forge to allow users to customize models around specific data, use cases, and requirements. Developers use Studio to build, govern, and improve agents and applications. Vibe, formerly known as Le Chat, functions as an AI chat assistant and coding agent for individuals and enterprises. Vibe integrates with 100+ tools including Slack, GitHub, and Jira through connectors and full MCP compatibility. The product includes Vibe for work and Vibe for code. Developers launch coding tasks in Vibe using remote agents or via the CLI in VS Code, JetBrains, or Zed. Vibe writes, tests, and deploys code across a codebase to allow developers to focus on architecture and business logic. The tool also handles document processing, contract reviews, and candidate screening with human checkpoints. Users upload spreadsheets or connect databases to spot patterns, flag anomalies, and extract signals. Vibe generates PDFs, board decks, strategic briefs, RFP responses, PRDs, reports, and more from research and data. Users can get briefed on missed information across email, chat, calendar, and internal knowledge in a single run. Users search the web, internal documents, and connected tools to compile analysis. If you rely on legacy code, Vibe moves codebases to new languages and architectures. Vibe handles bulk-updates for libraries and resolves breaking changes. Vibe also scaffolds multi-file features using tests, types, and documentation. The coding agent produces unit and integration tests. It writes code that matches team patterns. Vibe uses context from GitHub, GitLab, Jira, project boards, and business context to inform every coding decision. The free version of Vibe limits access to the best models.
European sovereignty and corporate investment
Mistral AI maintains its headquarters in Paris and employs over 1,000 people as of 2026. The company enables data residency in any cloud or datacenter. Users run models on-premises, in a private VPC, or on Mistral infrastructure. This flexibility keeps inference in-region and reserves dedicated GPU capacity. In September 2026, Samsung Electronics acquired a stake in Mistral in a transaction worth 3 billion Euro, which brought the company valuation to 21 billion Euro. This follows a 2 billion Euro investment in September 2025 that valued the firm at 12 billion Euro. ASML also holds a 11 percent stake after investing 1.5 billion USD. In June 2024, Mistral secured a 600 million Euro funding round. In May 2026, Mistral acquired the Linzer startup Emmi AI to expand its industrial AI and physics-based simulation capabilities. Mistral also expanded its platform in August 2026 to include GLM-5.2 from the Chinese company Z.ai. In Austria, the administration uses GovGPT, an application built on Mistral AI. Mistral’s founders, Arthur Mensch, Guillaume Lample, and Timothee Lacroix, met at the Ecole polytechnique. In February 2024, Microsoft announced that Mistral models would be available on the Microsoft Azure Cloud. Mistral’s founders previously worked at DeepMind and Meta Platforms. While Mistral provides high performance, can these European partnerships prevent total dependence on non-European hardware?




