Mistral AI Timeline: Model and Product Release History

Mistral AI is a French artificial intelligence company behind Mistral, Mixtral, Codestral, Voxtral, OCR, and other AI model families, along with products such as Vibe and Studio. Its first public model, Mistral 7B, arrived in September 2023.

This Mistral AI timeline tracks major model releases, product launches, availability changes, deprecations, and other release milestones from 2023 through September 2026.

Last updated: September 15, 2026.

Related: AI Model Release Calendar: Latest AI Model Releases by Date

Mistral AI timeline

2026

Date	Milestone
August 31, 2026	OCR 4.1 reached GA. Mistral moved OCR 4.1 to general availability after its July release.
August 20, 2026	Agentic Search released. Mistral introduced multi-step retrieval for finding, inspecting, navigating, and verifying information across indexed documents. Agentic Search became available through Search Toolkit and Libraries in Studio and Vibe.
August 4, 2026	Shieldstral 1.0 released in Public Preview. The 3.8B open-weight multimodal safety model evaluates text and images against natural-language policy questions. Mistral released the weights under Apache 2.0.
July 16, 2026	OCR 4.1 released. The update introduced block-level confidence scores and became the model used by the `mistral-ocr-latest` and `mistral-ocr-4` aliases.
July 8, 2026	Robostral Navigate released. Mistral introduced its first model for embodied navigation. The 8B model uses RGB camera input and natural-language instructions to guide robots through physical environments.
June 30, 2026	Leanstral 1.5 released in Public Preview. The Lean 4 model received updated training data and longer-context reasoning for formal proof engineering and automated theorem proving. Mistral scheduled this Labs release for retirement on September 30, 2026.
June 23, 2026	OCR 4 released. The document model introduced paragraph bounding boxes, structural block classification, confidence scores, and document parsing across 170 languages. It became available through the API and Document AI, with an enterprise self-hosting option.
May 28, 2026	Vibe launched as Mistral’s unified agent. Work handles longer productivity tasks, Code spans the CLI, VS Code, and remote coding sessions, and Chat preserves the earlier Le Chat experience.
May 28, 2026	Search Toolkit entered Public Preview. Mistral released an open-source framework for document ingestion, retrieval, and search evaluation across cloud, on-premises, and edge environments.
April 28, 2026	Mistral Medium 3.5 released. The 128B dense multimodal model introduced a 256,000-token context window, adjustable reasoning effort, coding capabilities, and agent features. Mistral released its weights under a Modified MIT license.
April 27, 2026	Workflows entered Public Preview. Studio gained durable multi-step processes for model calls, external APIs, tools, approvals, and work that can pause before resuming.
March 23, 2026	Voxtral TTS released. Mistral introduced multilingual text-to-speech generation, streaming audio, and zero-shot voice cloning.
March 17, 2026	Forge introduced. The enterprise system provides managed model training, alignment, evaluation, and deployment using an organization’s proprietary knowledge.
March 16, 2026	Mistral Small 4 released. The Apache 2.0 model unified instruction following, reasoning, coding, and agent capabilities in a 119B model with 6.5B active parameters and a 256,000-token context window.
March 16, 2026	Leanstral released in Labs. Mistral introduced its first open-weight model for Lean 4 formal proof engineering. Leanstral 1.5 replaced this release in June.
March 12, 2026	Mistral Moderation 2 released. The hosted moderation model expanded context to 128,000 tokens and added jailbreak detection for multilingual and multi-turn content.
February 4, 2026	Voxtral Mini Transcribe 2 and Voxtral Mini Transcribe Realtime released. The speech models introduced updated batch transcription and a dedicated real-time transcription model.
January 27, 2026	Vibe 2.0 released. The command-line coding agent update moved Devstral 2 from free access to paid API use and continued the transition from Mistral Code to the Vibe identity.

2025

Date	Milestone
December 18, 2025	OCR 3 released through the API. It introduced HTML table output, header and footer extraction, and hyperlink extraction. OCR 3 remains available for existing integrations and production workloads.
December 9, 2025	Devstral 2, Devstral Small 2, and Mistral Vibe released. Devstral’s second generation expanded Mistral’s software-engineering model family. Mistral Vibe became the new identity for its command-line coding agent.
December 2, 2025	Mistral 3 released. The generation included Mistral Large 3 and Ministral 3 models in 3B, 8B, and 14B sizes. The models introduced native multimodal capabilities and Apache 2.0 weights.
October 24, 2025	Mistral AI Studio announced in private beta. The developer platform introduced model prototyping, agents, knowledge, deployment, and observability under the Studio identity.
September 2, 2025	Le Chat gained custom MCP connectors and Memories. The product introduced connections to compatible services and an option to retain selected information across conversations.
August 12, 2025	Mistral Medium 3.1 released. The update advanced the Medium 3 multimodal model line before Medium 3.5 replaced it in 2026.
July 30, 2025	Codestral 25.08 released. Mistral updated its specialist code-completion model.
July 15, 2025	Voxtral released. Voxtral Small and Voxtral Mini introduced audio input for chat, and Voxtral Mini Transcribe introduced Mistral’s speech-to-text API. This was Mistral’s first public audio model family.
July 10, 2025	Devstral Medium and Devstral Small 1.1 released. The coding-agent family expanded beyond its original small model with a larger hosted model and an updated open version.
June 20, 2025	Mistral Small 3.2 released. API availability followed on June 23. The release updated the Small 3 family before Mistral Small 4 replaced it in 2026.
June 11, 2025	Mistral Compute announced. Mistral expanded into AI infrastructure designed for European and enterprise deployment requirements.
June 10, 2025	Magistral released. Magistral Small and Magistral Medium formed Mistral’s first dedicated reasoning model family. The Small model included downloadable weights, and Medium used hosted access.
June 4, 2025	Mistral Code introduced. The enterprise coding product included repository indexing, IDE integrations, code models, and deployment controls. Its command-line component later became Mistral Vibe.
May 28, 2025	Codestral Embed released. The model creates vector representations of source code for semantic search, retrieval, and repository indexing.
May 27, 2025	The new Agents API released. Mistral introduced managed agents with persistent conversations, tools, handoffs, and reusable instructions.
May 22, 2025	OCR 2 released. The second OCR generation introduced document annotations and structured information extraction through the API.
May 21, 2025	Devstral Small released. The open software-engineering model focused on codebase exploration, multi-file editing, and tool-using coding agents.
May 7, 2025	Mistral Medium 3 released. The multimodal model introduced a new Medium generation for enterprise and API workloads.
May 7, 2025	Le Chat Enterprise introduced. The business edition included organizational deployment, enterprise search, connectors, and administrative controls.
March 17, 2025	Mistral Small 3.1 released. The update introduced image understanding to the Small 3 family.
March 6, 2025	Mistral OCR released. Mistral’s first OCR model converted PDFs and images into structured text and tables through its document-understanding API.
February 17, 2025	Mistral Saba released. The regional model focused on Arabic and languages and cultural contexts associated with the Middle East and South Asia.
February 6, 2025	Le Chat relaunched for public use. The web and mobile assistant introduced document and image analysis, code execution, image generation, and news search.
January 30, 2025	Mistral Small 3 released. The 24B Apache 2.0 model began a new Small generation for local deployment, instruction following, and low-latency workloads.

2024

Date	Milestone
November 18, 2024	Mistral Large 2.1 and Pixtral Large released. Pixtral Large extended the Large generation into a 124B multimodal model. Le Chat gained web search, Canvas, document and image understanding, and image generation. Pixtral Large has since been deprecated.
November 6, 2024	Mistral Moderation and the Batch API released. The moderation model classified harmful text, and the Batch API introduced asynchronous processing for large workloads.
October 9, 2024	Ministral 3B and Ministral 8B released. The first Ministral generation focused on compact models for on-device and edge inference.
September 17, 2024	Mistral Small 2 released and Pixtral 12B reached La Plateforme. Mistral also introduced a free API tier, and Pixtral became available through Le Chat.
September 11, 2024	Pixtral 12B weights released. Mistral’s first public vision-language model combined text and image understanding in a 12B model with Apache 2.0 weights.
July 24, 2024	Mistral Large 2 released. The second Large generation improved multilingual text, code, reasoning, function calling, and long-context work. Mistral provided API access and released weights under the Mistral Research License.
July 18, 2024	Mistral NeMo released. Developed with NVIDIA, the 12B Apache 2.0 model used a 128,000-token context window and a new tokenizer for multilingual workloads.
July 16, 2024	Codestral Mamba and Mathstral released. Codestral Mamba explored a Mamba-based architecture for code, and Mathstral was a 7B model for mathematical reasoning.
June 5, 2024	The fine-tuning API released. La Plateforme introduced a managed process for customizing compatible Mistral models.
May 29, 2024	Codestral released. Mistral’s first dedicated code model handled code completion, fill-in-the-middle generation, and instruction-based coding tasks. Its license differed from the Apache 2.0 terms used for several earlier models.
April 17, 2024	Mixtral 8x22B released. The larger sparse mixture-of-experts model expanded the Mixtral family and became available as open weights and through La Plateforme.
February 26, 2024	Mistral Large, Mistral Small, and Le Chat released. Mistral Large became the company’s commercial flagship model. The launch also introduced a Small API model and the first public Le Chat product.

2023

Date	Milestone
December 11, 2023	Mixtral 8x7B released. The Apache 2.0 sparse mixture-of-experts model activates two of eight expert groups for each token. Mistral released base and instruction-tuned weights.
December 11, 2023	La Plateforme opened in beta. Mistral’s hosted developer platform introduced text-generation and embedding endpoints, including API access to Mistral 7B and Mixtral 8x7B.
September 27, 2023	Mistral 7B released. Mistral’s first public model was a 7.3B-parameter text model with base and instruction-tuned variants. The weights, reference implementation, and Apache 2.0 license permitted local and commercial deployment.
April 2023	Mistral AI founded. Arthur Mensch, Guillaume Lample, and Timothee Lacroix created the company in Paris.

Current Mistral AI models

Family	Current release	Role
Mistral Medium	Mistral Medium 3.5	Multimodal general-purpose model for reasoning, coding, tool use, and agent tasks.
Mistral Small	Mistral Small 4	Hybrid generalist model for instruction following, reasoning, coding, and agent workloads.
Mistral Large	Mistral Large 3	Large open-weight multimodal model.
Ministral	Ministral 3 3B, 8B, and 14B	Compact multimodal models for local and edge deployment.
Codestral	Codestral 25.08	Specialist model for code completion.
Leanstral	Leanstral 1.5	Lean 4 model for formal proof engineering and automated theorem proving.
Mistral OCR	OCR 4.1	Document model for text, layout, tables, structural blocks, and confidence scores.
Voxtral	Transcribe 2, Realtime, and TTS	Models for transcription, live transcription, and speech generation.
Mistral Moderation	Mistral Moderation 2	Hosted text moderation model with jailbreak detection.
Shieldstral	Shieldstral 1.0	Open-weight multimodal safety model for text and image classification.
Robostral	Robostral Navigate	Model for robot navigation using camera input and natural-language instructions.

Current Mistral products

Mistral organizes its product platform into Vibe, Studio, and Admin. Vibe includes Work, Code, and Chat modes. Studio contains model access, application development, agents, workflows, connectors, and deployment features. Admin handles organization, workspace, billing, identity, and access controls.

Historical Mistral model families

Model family	Historical role
Mixtral	Sparse mixture-of-experts family introduced with Mixtral 8x7B in 2023 and expanded with Mixtral 8x22B in 2024.
Pixtral	Mistral’s first vision-language family, beginning with Pixtral 12B and followed by Pixtral Large. Multimodal capabilities later moved into newer general-purpose models.
Magistral	Dedicated reasoning family introduced in 2025 with Small and Medium models. Its standalone reasoning models were later deprecated.
Devstral	Software-engineering family for repository exploration, multi-file editing, and coding agents. Devstral 2 was later deprecated in favor of newer general-purpose models.
Mistral NeMo	A 12B model developed with NVIDIA in 2024 with a 128K context window and multilingual tokenizer.
Mathstral	A 7B specialist model released in 2024 for mathematical reasoning.
Codestral Mamba	An experimental code model based on the Mamba architecture, released in 2024.
Mistral Saba	A regional model introduced in 2025 for Arabic and languages and cultural contexts associated with the Middle East and South Asia.

Mistral product name changes

Earlier product	Current product	Change
La Plateforme	Studio	Mistral introduced the Studio identity in October 2025 and expanded it into the developer environment for models, agents, workflows, connectors, and deployment.
Le Chat	Vibe Chat	Le Chat became the Chat mode within the unified Vibe product in May 2026.
Mistral Code / Mistral Vibe CLI	Vibe Code	Mistral’s coding product now spans the terminal, VS Code, and remote coding sessions under Vibe Code.

FAQs

What are the newest Mistral AI models?

As of September 15, 2026, Mistral’s general-purpose lineup includes Mistral Medium 3.5, Mistral Small 4, Mistral Large 3, and Ministral 3. Recent specialist releases include Shieldstral 1.0, OCR 4.1, Leanstral 1.5, Robostral Navigate, Voxtral TTS, and Voxtral Mini Transcribe 2.

Are all Mistral AI models open source?

No. Mistral uses several licensing and access models. Some weights use Apache 2.0 or Modified MIT licenses, some releases use research or other licenses, and some models are available only as hosted services. Usage rights depend on the license attached to each release.

What is the difference between Le Chat and Vibe?

Le Chat was Mistral’s original assistant product. Vibe is the unified product with Work, Code, and Chat modes. Chat preserves the Le Chat experience, Work handles longer productivity tasks, and Code operates across terminal, VS Code, and remote coding sessions.