Mistral AI investigated a vLLM memory leak in a disaggregated Prefill/Decode deployment of Mistral Medium 3.1. System memory grew linearly at 400 MB per minute, causing out-of-memory failures within hours. It occurred only on the decode side with graph compilation and KVCache transfer via NIXL. Heaptrack showed stable heap memory; the issue was outside the heap.
An encoder (optical system) maps objects to noiseless images, which noise corrupts into measurements. Our information estimator uses only these noisy measurements and a noise model to quantify how well measurements distinguish objects. Many imaging systems produce measurements that humans never see or cannot interpret directly. Your smartphone processes raw sensor data through algorithms before producing the final photo.
Mistral AI released Mistral OCR 3 with an overall 74% win rate against Mistral OCR 2 on forms, scans, complex tables and handwriting. The company says its accuracy exceeds enterprise document-processing and AI-native OCR solutions.
Mistral AI released the next-generation Devstral 2 coding family, including 123B Devstral 2 under a modified MIT licence and 24B Devstral Small 2 under Apache 2.0, both open-source.
Mistral released the Mistral 3 family, comprising 14B, 8B and 3B small dense models and its strongest yet Mistral Large 3, a sparse MoE with 41B active and 675B total parameters. All are open-source under Apache 2.0.
We are building a multiyear partnership with SAP to deliver a fully sovereign AI stack for Germany and Europe, integrating our models into SAP’s AI Foundation, and co-developing industry-specific solutions for Europe’s most complex industries and administrations. With Helsing, we are accelerating the development of vision-language-action models for real-world defense and security applications.
In this post, I’ll introduce a reinforcement learning (RL) algorithm based on an “alternative” paradigm: divide and conquer . Unlike traditional methods, this algorithm is not based on temporal difference (TD) learning (which has scalability challenges ), and scales well to long-horizon tasks. We can do Reinforcement Learning (RL) based on divide and conquer, instead of temporal difference (TD) learning.
Mistral has released Mistral AI Studio, a production-grade AI platform for enterprise teams built on three pillars: observability, Agent Runtime and AI Registry.
We are announcing a Series C funding round of 1.7B€ at a 11.7B€ post-money valuation. This investment fuels our scientific research to keep pushing the frontier of AI to tackle the most critical and sophisticated technological challenges faced by strategic industries. The Series C funding round is led by leading semiconductor equipment manufacturer, ASML Holding NV (ASML).
Le Chat Memories combines automatic saving, intelligent recall and visible sources. Users can disable memory, start memory-free incognito chats, edit or delete individual entries, export and import from other platforms. Memory Insights highlights trends and summaries from personal data, with categorisation and instant forgetting planned.
Mistral launched a beta directory of more than 20 secure MCP-based connectors in Le Chat across data, productivity, development, automation and business. It connects tools including Databricks, Snowflake, GitHub, Atlassian, Asana, Outlook, Box, Stripe and Zapier, and supports custom connectors or any remote MCP server.
What exactly does word2vec learn, and how? Answering this question amounts to understanding representation learning in a minimal yet interesting language modeling task. Despite the fact that word2vec is a well-known precursor to modern language models, for many years, researchers lacked a quantitative and predictive theory describing its learning process.
Mistral fine-tuned Pixtral-12B with LoRA, substantially outperforming the untuned baseline on Aerial Image Dataset (AID) satellite classification and reducing hallucinated invalid class names. Fine-tuning uses Mistral's API or LaPlateforme UI without extensive hyperparameter tuning, with 8,000 training and 2,000 test samples.
Mistral AI released Codestral 25.08 and a complete enterprise coding stack comprising Codestral, Codestral Embed, Devstral and the Mistral Code IDE plugin.
Mistral AI, Carbone 4 and France's ADEME completed the first full-lifecycle AI model analysis. By January 2025, Large 2 training and 18 months of use produced 20.4 ktCO₂e, consumed 281,000 cubic metres of water and 660 kg Sb eq of resources.
Mistral added features to Le Chat including preview Deep Research, Voxtral-powered voice mode, multilingual reasoning with Magistral, Projects for organising conversations and image editing with Black Forest Labs.
Mistral AI released Voxtral speech-understanding models in 24B and 3B versions, both under Apache 2.0 and available via its API. With a 32k-token context, they handle up to 30 minutes of transcription or 40 minutes of audio understanding, including built-in question answering and summaries, automatic multilingual detection and voice-triggered function calls, inheriting Mistral Small 3.1's text capabilities.
Mistral AI partnered with All Hands AI to release Devstral Medium and upgrade Devstral Small 1.1.
Mistral AI launched AI for Citizens, helping governments and public institutions strategically apply AI to transform public services, innovate and safeguard competitiveness. It offers open-source models, self-hosting, data sovereignty and tailored R&D, with government and public-sector partners in France, Luxembourg, Singapore, the Netherlands, the UK and Switzerland.
Mistral AI launched Mistral Compute, a private integrated AI infrastructure stack covering GPUs, orchestration, APIs, products and services, ranging from bare-metal servers to fully managed PaaS.
Mistral AI released its first reasoning model, Magistral, comprising the 24B open-source Magistral Small and enterprise-focused Magistral Medium.
Mistral released Mistral Code, an AI coding assistant combining Codestral, Codestral Embed, Devstral and Mistral Medium. It supports cloud, dedicated-capacity and local air-gapped GPU deployment, keeping code within enterprise boundaries.
Mistral AI released its first code-focused embedding model, Codestral Embed, outperforming Voyage Code 3, Cohere Embed v4.0 and OpenAI's large embedding model on real-world code retrieval.
Mistral AI released Agents API, combining its language models with built-in connectors for code execution, image generation, document libraries and web search, plus MCP tools, persistent memory across conversations and agent orchestration.
Mistral AI and All Hands AI jointly released Devstral, an agentic LLM for software engineering tasks, open-source under Apache 2.0.
Mistral AI released Mistral Medium 3, claiming at least 90% of Claude Sonnet 3.7’s performance on several benchmarks. Pricing is $0.40 per million input tokens and $2 per million output tokens.
Mistral released enterprise AI assistant Le Chat Enterprise, powered by the new Mistral Medium 3. It offers enterprise search, agent building, custom data and tool connectors, document libraries, custom models and hybrid deployment, with all features rolling out over the next two weeks.
Mistral proposes judge LLMs scoring generator responses on numerical, binary or qualitative scales, then taking weighted averages across evaluation datasets to assess RAG systems.
Mistral AI released Mistral Small 3.1, improving text performance and multimodal understanding over Small 3, expanding context to 128k tokens and reaching 150 tokens per second, under Apache 2.0.
Mistral AI introduced Mistral OCR, an optical character recognition API that reads images and PDFs and outputs text and images interleaved in order. It is already the default document understanding model for millions of Le Chat users.
Mistral AI introduced TranscriptToPRDTicket, comprising PRDAgent and TicketCreationAgent powered by Mistral Large 2. It turns meeting transcripts into PRDs and development tickets, automatically creating them in tools such as Linear and Jira.
Mistral Saba is a 24B parameter model trained on meticulously curated datasets from across the Middle East and South Asia. The model provides more accurate and relevant responses than models that are over 5 times its size, while being significantly faster and lower cost.
Today, we unveil the latest le Chat, a comprehensive AI assistant for your life and work. We're also introducing Pro and Team tiers of service, as well as an Enterprise tier in private preview. And starting today, le Chat is available on iOS, Android, and soon, on private infrastructure for businesses.
Today we’re introducing Mistral Small 3, a latency-optimized 24B-parameter model released under the Apache 2.0 license.