Mistral Large 4 (mistral-large-4) Public Preview
Mistral Large 4 is an open-weight, general-purpose multimodal model with a Mixture-of-Experts architecture and a 1M context window. Open weights are coming soon. Launch pricing: 50% off for 2 weeks.
Last updated · 87 releases tracked since September 2023
Mistral AI news, updates and changelog entries covering Mistral models such as Mistral Small and Voxtral, plus the Mistral API and platform tools. Entries come from Mistral AI announcements and the Mistral Docs changelog, each dated and linked to its source.
Mistral AI released Magistral Medium 1.2 and Magistral Small 1.2, OCR 4.1 GA, and Z.ai GLM 5.3 GA in September 2026. The OCR API introduced an include_blocks parameter for block-level data, and several older models were deprecated.
Summary of September 2026, generated from the releases listed below.
This month
1
-8 vs last month
Last 12 months
52
Busiest: March 2026
Most shipped
Model
30 releases
Newest first. Each entry links to the official page it came from.
Mistral Large 4 is an open-weight, general-purpose multimodal model with a Mixture-of-Experts architecture and a 1M context window. Open weights are coming soon. Launch pricing: 50% off for 2 weeks.
OCR 4.0 is deprecated and retires on September 30, 2026. Use OCR 4.1 instead at the same price.
Z.ai GLM 5.2 ( zai-glm-5-2 ) is deprecated and retires on October 31, 2026. Use Z.ai GLM 5.3 ( zai-glm-5-3 ) instead, at the same price.
OCR 4.0 ( mistral-ocr-4-0 ) is deprecated and retires on September 30, 2026. Use OCR 4.1 ( mistral-ocr-4-1 or mistral-ocr-latest ) instead, at the same price.
Leanstral 1.5 ( labs-leanstral-1-5 ), an experimental Labs model, is deprecated and retires on September 30, 2026.
Z.ai GLM 5.3 ( zai-glm-5-3 ) is now Generally Available.
Introducing include_blocks parameter in OCR API for block-level bounding boxes and structural labels.
We released Magistral Medium 1.2 ( magistral-medium-2509 ) and Magistral Small 1.2 ( magistral-small-2509 ).
OCR 4.1 ( mistral-ocr-4-1 ) is now Generally Available.
We released Mistral Medium 3.1 (mistral-medium-2508).
Added a new parameter p to the chunks streamed back by the Completion API. SECURITY API UPDATED
We released Codestral 2508 ( codestral-2508 ).
We released Magistral Medium 1.1 ( magistral-medium-2507 ) and Magistral Small 1.1 ( magistral-small-2507 ).
We released a Document Library API to manage libraries.
We released OCR 4.1 ( mistral-ocr-4-1 ). mistral-ocr-latest and mistral-ocr-4 now point to it.
Leanstral 1.5 is a free Apache-2.0 licensed model with 6B active parameters that delivers major performance upgrades in formal verification, solving 587/672 PutnamBench problems and achieving state-of-the-art results on FATE-H (87%) and FATE-X (34%). It excels in agentic proof e…
The OCR API confidence_scores_granularity parameter now supports "block" granularity. "page" returns page-level scores only, "block" returns page-level and block-level scores, and "word" returns page-level and word-level scores.
We released Leanstral 1.5 ( labs-leanstral-1-5 ), an updated Lean 4 formal proof engineering model with improved SFT mixture quality and extended long-context reasoning. This model will be retired on September 30, 2026 .
The pages parameter in our OCR API now also accepts a string of comma-separated digits and ranges (e.g. "0,1,2" , "0-5" , or "0,2-4" ) in addition to a list of integers.
Introducing include_blocks in our OCR API. When set to true , each page returns a blocks array with paragraph-level bounding boxes and a structural label ( text , title , list , table , image , equation , caption , code , references , aside_text , header , footer , signature ) i…
We released OCR 4 ( mistral-ocr-4-0 ). mistral-ocr-latest now points to it.
We launched Vibe, our unified agent at chat.mistral.ai , available in three modes: Work for productivity on web and mobile, with Skills, Workflows, Connectors, Libraries, and scheduled tasks. Code for developers, covering the Vibe CLI , the VS Code extension , and Vibe Code Web…
Search Toolkit is a composable framework for building production search pipelines for AI applications, offering ingestion, retrieval, and evaluation in a single framework with a shared interface.
We released Mistral Medium 3.5 ( mistral-medium-3-5 ), our frontier-class multimodal model optimized for agentic and coding use cases, with adjustable reasoning via the reasoning_effort parameter. Released as open weights under a Modified MIT license.
We released Voxtral TTS ( voxtral-tts-2603 ), our state-of-the-art text-to-speech model with zero-shot voice cloning, multilingual support, and real-time streaming.
Mistral AI has launched Voxtral TTS, a 4B-parameter text-to-speech model delivering state-of-the-art multilingual voice generation with realistic, emotionally expressive speech in 9 languages, low latency, and easy voice adaptation. The model excels in contextual understanding,…
Mistral released Voxtral TTS, a text-to-speech model with zero-shot voice cloning, multilingual support, and real-time streaming.
We released Mistral Small 4 ( mistral-small-2603 ), a hybrid model unifying instruct, reasoning, and coding in a single multimodal model with a 256k context window.
We released Leanstral ( labs-leanstral-2603 ), our first open-source code agent designed for Lean 4 formal proof engineering.
Introducing Forge, a system for enterprises to build AI models trained on proprietary knowledge, enhancing control and customization
Mistral Small 4 is a hybrid model unifying instruct, reasoning, and coding in a single multimodal model with a 256k context window.
Mistral released Small 4, a hybrid multimodal model that unifies instruct, reasoning, and coding with a 256k context window.
Open-source code agent designed for Lean 4, optimized for proof engineering tasks with efficient sparse architecture and parallel inference capabilities.
Mistral released Leanstral, its first open-source code agent for Lean 4 formal proof engineering.
We released Mistral Moderation 2603 ( mistral-moderation-2603 ).
Mistral released Moderation 2603 and added custom guardrails support for Agents and Conversations, tightening moderation and policy controls in the API.
We added Custom Guardrails support for Agents and Conversations. Guardrails can now be configured directly on an Agent via the guardrails parameter. Guardrails can be passed per-request on POST /v1/conversations using the guardrails field. Guardrails can be passed per-request on…
Mistral released Voxtral Mini Transcribe 2 and Voxtral Realtime, adding diarization, context biasing, and low-latency transcription with a new Studio audio playground.
We released Voxtral Mini Transcribe 2 ( voxtral-mini-2602 ) and Voxtral Mini Transcribe Realtime ( voxtral-mini-transcribe-realtime-2602 ).
Introducing diarize in our Audio Transcriptions API.
Introducing context biasing in our Audio Transcriptions API.
Vibe 2.0 with enhanced features and documentation restructuring.
Mistral Vibe 2.0 upgrades the terminal-native coding agent with custom subagents, clarification prompts, slash-command skills, unified agent modes, and new plan access.
Document Annotations update and improvements, 8 page limit removed.
Devstral 2.0 moves to paid API access.
Mistral OCR 3 introduces breakthrough performance with 74% overall win rate over Mistral OCR 2 on forms, scanned documents, complex tables, and handwriting. It offers state-of-the-art accuracy, powers Document AI Playground for parsing PDFs/images into text/JSON, and is availabl…
Mistral Small Creative is a Labs model released as a creative variant.
Mistral Large 3 and Ministral 3 models released with new variants.
Mistral Large 2.1 and Pixtral Large models released with enhanced capabilities.
Announces Codestral 25.08 with 30% increase in accepted completions, 10% more retained code, and 50% fewer runaway generations for enterprise coding workflows.
Mistral Large 2 model released with improved performance.
Mistral Nemo is an open-weight model for general-purpose tasks.
Codestral Mamba and Mathstral models released for coding and mathematical tasks.
State-of-the-art speech understanding models available in 24B and 3B variants under Apache 2.0 license, with API access and open-source deployment options.
Announcing Magistral - the first reasoning model by Mistral AI - excelling in domain-specific, transparent, and multilingual reasoning.
AI-powered coding assistant with in-IDE support, local deployment, and enterprise tooling
Codestral Latest (codestral-2405) model released with new capabilities.
Introducing Devstral, an agentic LLM for software engineering tasks built in collaboration with All Hands AI. Outperforms open-source models on SWE-Bench Verified benchmark and released under Apache 2.0 license.
Mistral Medium 3 introduces a new class of models balancing SOTA performance with 8X lower cost and simpler deployability for enterprise use cases.
Open-mixtral-8x22b model released with new variants.
Mistral Small 3.1 model released with improved performance.
Mistral Small 3.1 is an improved open-source model with enhanced text performance, multimodal understanding, and a 128k token context window. It outperforms models like Gemma 3 and GPT-4o Mini while maintaining fast inference speeds of 150 tokens/sec under Apache 2.0 license.
Mistral OCR (mistral-ocr-2503) and document understanding features released.
Mistral OCR is an Optical Character Recognition API that sets a new standard in document understanding. It comprehends each element of documents-media, text, tables, equations-with unprecedented accuracy and cognition.
Mistral Saba model released for specific use cases.
Mistral Small 3 model released with enhanced features.
Codestral 2501 model released for coding tasks.
Codestral 25.01 is an upgraded coding model with a more efficient architecture and improved tokenizer, generating code up to 2x faster. It leads in FIM use cases and demonstrates significant performance improvements across multiple benchmarks.
Pixtral Large is a 124B open-weights multimodal model built on Mistral Large 2, featuring frontier-level image understanding and competitive performance on multimodal benchmarks.
Mistral Moderation API provides content moderation capabilities for text and conversational content, enabling users to customize safety standards and deploy scalable moderation solutions.
Introducing Ministral 3B and 8B, the world's best edge models for on-device computing and at-the-edge use cases. These sub-10B models offer advanced knowledge, reasoning, function-calling, and efficiency with up to 128k context length (32k on vLLM) and specialized attention patt…
Pixtral 12B is a multimodal model trained with interleaved image and text data, offering strong performance on multimodal tasks and text-only benchmarks. It features a 400M parameter vision encoder and a 12B parameter multimodal decoder, supporting variable image sizes and multi…
Released free tier on la Plateforme, updated pricing across models, Mistral Small v24.09 with improved capabilities, and Pixtral 12B vision support on le Chat.
Released mistralai 1.0 for Python and Typescript with improved usability and consistency
Mistral Large 2 is a new model with a 128k context window, supports multiple languages, and is available on la Plateforme.
Released a 12B model with 128k context window, state-of-the-art reasoning, world knowledge, and coding accuracy. Developed in collaboration with NVIDIA, compatible with Mistral 7B systems, and available under Apache 2.0 license.
Mathstral is a specialized model for advanced mathematical problems requiring complex, multi-step logical reasoning. It achieves state-of-the-art reasoning capacities in its size category, with 56.6% on MATH and 63.47% on MMLU benchmarks. Weights are hosted on HuggingFace.
Codestral Mamba is a new model using Mamba architecture with linear time inference, designed for code productivity. It offers efficient handling of long sequences and is available under Apache 2.0 license.
Codestral is a 22B open-weight generative AI model designed for code generation tasks, supporting 80+ programming languages and available via HuggingFace and Mistral AI endpoints.
Mixtral 8x22B is a sparse Mixture-of-Experts (SMoE) model with 39B active parameters out of 141B, offering cost efficiency. It is fluent in multiple languages, has strong math/coding capabilities, and 64K tokens context window. Released under Apache 2.0 license.
Mistral Large is our new cutting-edge text generation model with top-tier reasoning capabilities, available through la Plateforme and Azure. It excels in complex multilingual reasoning tasks, code generation, and achieves strong results on benchmarks, ranking second globally aft…
High-quality sparse mixture of experts model (SMoE) with open weights, outperforming Llama 2 70B with 6x faster inference. Licensed under Apache 2.0.
Mistral AI launches beta access to their platform services, including three chat endpoints (mistral-tiny, mistral-small, mistral-medium) and an embedding endpoint (mistral-embed), with varying performance/price tradeoffs and API specifications.
Mistral 7B is a 7.3B parameter model that outperforms Llama 2 13B on all benchmarks and approaches CodeLlama 7B performance on code while remaining good at English tasks. It uses Grouped-query attention (GQA) for faster inference and Sliding Window Attention (SWA) to handle long…
Releases per month over the last 12 months. Busiest month: March 2026, 13 releases.
Releases in the last 12 months by category and product.
The most recent Mistral AI update we track is “Mistral Large 4 (mistral-large-4) Public Preview”, released on October 6, 2026.
Mistral AI shipped 1 update in October 2026, including “Mistral Large 4 (mistral-large-4) Public Preview”.
We tracked 52 updates from Mistral AI in the last 12 months, about 4 a month. The busiest month was March 2026, with 13 releases.
Mistral AI publishes release notes on Mistral Docs Changelog, linked at the top of this page. Every entry here links back to the official page it came from.
Newsletter
This viewscreen turns the calendar into a recurring digest: the sharpest launches, model drops, and product moves delivered in one clear read.
The standout launches, model drops, and platform moves worth tracking.
One tight read instead of bouncing between company blogs and timelines.
Clear summaries you can scan fast and forward to the team when needed.
Monthly dispatch
Get a clean monthly dispatch covering the launches that mattered, not every minor product breadcrumb.
One email each month, tuned for signal over noise.