Mistral AI

Mistral AI is a French artificial intelligence company headquartered in Paris that develops and deploys foundation models and related products for text, programming code, images, speech, document processing, autonomous agents, formal reasoning, physics, and robotics. It was founded in April 2023 by Arthur Mensch, Guillaume Lample, and Timothée Lacroix.

The company follows a mixed distribution strategy. It releases the weights of many models, often under the Apache License 2.0, while other checkpoints use the Mistral Research License, Mistral AI Non-Production License, a modified MIT license, or are distributed only as commercial services. Consequently, Mistral AI's complete model portfolio should not be described as uniformly "open source"; open-weight is the more precise general term for downloadable models whose complete training data and end-to-end training pipeline have not been released.

Mistral AI became known for combining established efficiency techniques such as Grouped-Query Attention (GQA), Sliding Window Attention (SWA), and sparse Mixture-of-Experts architectures. The techniques themselves predate the company; Mistral's contribution was their implementation and combination in models such as Mistral 7B and Mixtral.

As of July 11, 2026, Mistral AI's featured models include Mistral Medium 3.5, Mistral Small 4, Voxtral Mini Transcribe Realtime, Voxtral TTS, Mistral OCR 4, and Devstral 2. On July 8, 2026, the company also introduced Robostral Navigate, its first model for embodied robotic navigation.

History

Mistral AI was created in April 2023 by three French artificial-intelligence researchers. Arthur Mensch, who became chief executive officer, had previously worked at Google DeepMind. Guillaume Lample, now chief science officer, and Timothée Lacroix, now chief technology officer, had worked at Meta's artificial-intelligence research organization.

The company states that its mission is to make frontier artificial intelligence more accessible and customizable while giving organizations greater control over deployment and data. In practice, this strategy combines downloadable model weights, hosted APIs, enterprise software, customization services, and infrastructure products rather than relying exclusively on open releases.

The company attracted several large financing rounds:

• June 2023: Mistral AI raised €105 million in a seed round led by Lightspeed Venture Partners, approximately one month after beginning operations.
• December 2023: The company raised €385 million in a Series A round. Mistral AI did not disclose a valuation, but a source cited by Reuters estimated it at approximately €2 billion.
• February 2024: Microsoft and Mistral AI announced a distribution partnership through Microsoft Azure. Microsoft provided a €15 million convertible investment, equivalent to approximately US$16 million at the time. In May 2024, the United Kingdom's Competition and Markets Authority concluded that the arrangement did not give Microsoft material influence over Mistral AI and therefore did not qualify for a merger investigation.
• June 2024: A €600 million Series B round valued the company at approximately €5.8 billion. The proceeds included both equity and debt financing.
• September 2025: Mistral AI raised €1.7 billion in a Series C round at a post-money valuation of €11.7 billion. The round was led by semiconductor-equipment manufacturer ASML, which invested €1.3 billion and became Mistral AI's largest shareholder with a stake of approximately 11%.
• March 2026: The company raised US$830 million in debt financing to purchase 13,800 Nvidia accelerators for a data centre near Paris. This was debt financing for infrastructure and should not be treated as a new equity valuation round.

In February 2026, Mistral AI announced an agreement to acquire French serverless-cloud company Koyeb, its first corporate acquisition. In May, it acquired Austrian physics-modeling startup Emmi AI to expand into industrial simulation, engineering, and physical AI.

The July 2026 launch of Robostral Navigate marked the company's entry into embodied navigation. The 8-billion-parameter model accepts images from a single RGB camera and natural-language navigation instructions. It does not require LiDAR or depth sensors, but its announced capability is navigation rather than general robotic manipulation.

Technical Architectural Features

Mistral AI does not use a single architecture across all of its models. Its portfolio includes dense decoder-only transformers, sparse MoE transformers, multimodal models with vision encoders, speech systems, document-processing models, and state-space models.

Transformer with Enhancements (Mistral 7B)

The company's first model, Mistral 7B, is a 7.3-billion-parameter decoder-only transformer with 32 layers, a hidden dimension of 4,096, 32 query attention heads, and eight key-value heads.

Its principal efficiency mechanisms were:

• Sliding Window Attention (SWA): Each token attends directly to a fixed window of up to 4,096 preceding tokens instead of every earlier token. With a fixed window size, the attention cost grows linearly with sequence length rather than quadratically. Information can propagate beyond one window through successive transformer layers, but this theoretical receptive field does not guarantee reliable performance at arbitrary sequence lengths.
• Grouped-Query Attention (GQA): Mistral 7B uses 32 query heads and eight key-value heads, equivalent to four query heads sharing each key-value head. The frequently stated ratio of 8:1 is therefore incorrect for this model. GQA reduces the size of the KV cache and the memory bandwidth required during autoregressive generation.
• Context Length: The current official model card lists a supported context of 8,000 tokens. The original configuration contains a max_position_embeddings value of 32,768, but that configuration field should not be treated by itself as proof of a validated 32K context. The official 8K figure is therefore used for the released model.

Mistral 7B was retired from Mistral's supported hosted-model lifecycle on March 30, 2025. Its Apache-2.0-licensed weights remain available for self-hosting; retirement of a hosted endpoint does not revoke the weight license.

Sparse Mixture-of-Experts (MoE)

The Mixtral series replaces each dense feed-forward sublayer with multiple parallel expert networks. A router determines which experts process each token.

In Mixtral 8x7B, every MoE layer contains eight feed-forward experts, with two selected for each token. The model contains approximately 46.7 billion total parameters and activates approximately 12.9 billion parameters per token. In Mixtral 8x22B, approximately 39 billion of 141 billion total parameters are active per token.

Sparse activation reduces arithmetic per token relative to a dense model with the same total parameter count. It does not reduce storage requirements to the active-parameter count: all expert weights generally have to remain stored or accessible, and real inference speed also depends on memory bandwidth, routing overhead, expert parallelism, batch size, quantization, and hardware utilization.

Mixtral 8x7B uses a fully dense 32K attention context rather than the Sliding Window Attention configuration of Mistral 7B. Mixtral 8x22B supports a 64K context.

Mistral later returned to large sparse models with:

• Mistral Large 3 — 675 billion total parameters and 41 billion active parameters, released under Apache 2.0 in December 2025.
• Mistral Small 4 — 119 billion total parameters, approximately 6 billion active parameters in its transformer blocks, 128 experts, and four selected experts per token. The model supports text and image inputs and a 256K context window.
• Leanstral 1.5 — 119 billion total parameters and approximately 6 billion active parameters, specialized for formal proofs and code verification in Lean 4.

Mamba Architecture (SSM)

In July 2024, Mistral AI released Codestral Mamba, a 7.285-billion-parameter coding model based on the Mamba 2 selective state-space model architecture. Unlike transformer attention, Mamba updates a recurrent state and can perform inference with computational cost that scales linearly with sequence length.

Mistral described Mamba models as theoretically able to model sequences of unbounded length. This does not mean that Codestral Mamba provides a practically unlimited or perfectly retained context: Mistral tested its in-context retrieval up to 256,000 tokens, which is also the context length listed in the official model card. Quality beyond the trained and tested range depends on the task, implementation, numerical precision, and information retained in the recurrent state.

The model was retired from Mistral's supported hosted lifecycle on June 6, 2025, although its Apache-2.0-licensed weights remain available.

Chronology and Models

Major Mistral AI Model Releases

Month / Year	Model	Parameters (billion)	Key Features	License
09 / 2023	Mistral 7B	7.3	Dense decoder-only transformer with GQA and 4,096-token SWA; officially supported 8K context. Retired from the hosted lifecycle in March 2025.	Apache 2.0
12 / 2023	Mixtral 8x7B	46.7 total / 12.9 active	Eight feed-forward experts with top-2 routing; full-attention 32K context; multilingual and coding capabilities.	Apache 2.0
02 / 2024	Mistral Small 1 / Mistral Large 1	Not disclosed	Hosted commercial text models. Mistral Large 1 supported a 32K context and was distributed through Mistral's platform and Microsoft Azure.	Commercial / access to weights by agreement
04 / 2024	Mixtral 8x22B	141 total / 39 active	Sparse top-2 MoE model with a 64K context window and native function calling.	Apache 2.0
05 / 2024	Codestral 22B	22	Code-completion and code-generation model supporting more than 80 programming languages and a 32K context.	Mistral AI Non-Production License; separate commercial license for production self-deployment
07 / 2024	Codestral Mamba / Mathstral	7.3 / 7	Mamba 2 code model tested to 256K; specialized mathematical reasoning model with a 32K context.	Apache 2.0
07 / 2024	Mistral NeMo	12	Dense multilingual model developed with Nvidia; 128K context and the Tekken tokenizer.	Apache 2.0
07 / 2024	Mistral Large 2	123	Dense multilingual model with a 128K context, function calling, and support for more than 80 programming languages.	Mistral Research License for research and non-commercial use; separate commercial license
09 / 2024	Pixtral 12B	12B decoder + 0.4B vision encoder	Native image-and-text model supporting variable image resolution, multiple images, and a 128K context. Later marked deprecated.	Apache 2.0
01–03 / 2025	Mistral Small 3 / 3.1	24	Small 3 focused on low-latency text generation; Small 3.1 added image understanding and expanded the context window to 128K.	Apache 2.0
03 / 2025	Mistral OCR	Not disclosed	Document-understanding service for extracting text, tables, equations, and images from PDFs and other document formats.	Commercial service
05 / 2025	Mistral Medium 3 / Devstral Small 1	Medium: not disclosed; Devstral: 24	Medium 3 was a hosted multimodal enterprise model. Devstral was an open-weight model for multi-file software-engineering agents.	Medium 3: proprietary; Devstral: Apache 2.0
06 / 2025	Magistral Small / Medium	Small: 24; Medium: not disclosed	First Mistral reasoning-model family, with native multilingual reasoning and increased inference-time computation.	Small: Apache 2.0; Medium: commercial API
07 / 2025	Voxtral Mini / Small	3 / 24	Speech-understanding models for transcription, question answering, summarization, translation, and voice-initiated function calling; 32K context.	Apache 2.0
12 / 2025	Mistral 3	Large 3: 675 total / 41 active; Ministral 3: 3, 8, and 14	Mistral Large 3 is a multimodal sparse MoE model. Ministral 3 consists of dense edge-oriented base, instruction, and reasoning models with image understanding.	Apache 2.0
12 / 2025	Devstral 2 / Devstral Small 2	123 / 24	Agentic coding models with 256K contexts. Mistral reported 72.2% and 68.0%, respectively, on SWE-bench Verified under its evaluation setup.	Devstral 2: Modified MIT; Devstral Small 2: Apache 2.0
03 / 2026	Mistral Small 4	119 total / 6 active	Sparse multimodal model combining instruction following, configurable reasoning, and agentic coding; 128 experts, four selected per token, and a 256K context.	Apache 2.0
05 / 2026	Mistral Medium 3.5	128	Dense multimodal model with a 256K context and configurable reasoning, optimized for long-running coding and productivity agents.	Modified MIT with a revenue-based restriction
06 / 2026	Mistral OCR 4	Not disclosed	Document-intelligence model with bounding boxes, block classification, confidence scores, support for 170 languages, API access, and enterprise self-hosting.	Commercial API and enterprise self-hosting terms
07 / 2026	Leanstral 1.5	119 total / 6 active	Sparse model for theorem proving and software verification in Lean 4.	Apache 2.0
07 / 2026	Robostral Navigate	8	Embodied-navigation model using natural-language instructions and images from one RGB camera; trained in simulation.	Public license not stated in the release

The table lists major releases rather than every model version. Mistral has also released updated Codestral, Magistral, OCR, moderation, embedding, reranking, audio-transcription, and text-to-speech checkpoints. Hosted aliases and model lifecycle status can change independently of downloadable-weight availability.

Comparison with Competitors

Performance comparisons involving Mistral models are dependent on the checkpoint, prompt template, evaluation harness, sampling parameters, context configuration, tool access, inference-time reasoning budget, and hardware. Most comparisons in Mistral release announcements are produced by the model developer and should not be treated as independent proof of general superiority.

• vs. Llama (Meta): In Mistral's 2023 evaluation, Mistral 7B scored above Llama 2 13B on all benchmarks included in the technical report and approached or exceeded Llama 2 34B on several tasks. Mixtral 8x7B later outperformed Llama 2 70B on most of the evaluations in Mistral's report while activating fewer parameters per token. These are historical, developer-reported comparisons with models available in 2023; they do not imply that every Mistral model is superior to every subsequent Llama model.
• vs. GPT (OpenAI): The Mixtral report found Mixtral 8x7B comparable to or stronger than the tested GPT-3.5 endpoint on most listed benchmarks. It did not claim equivalence to GPT-4 across all tasks. Later Mistral models have been compared with GPT-4o and other proprietary systems on selected benchmarks, but model versions and evaluation conditions differ.
• vs. Claude (Anthropic): Mistral's Large 2 announcement described the model as competitive with Claude 3 Opus on selected reasoning and coding evaluations. This does not support an unrestricted statement that a Mistral model "surpasses Claude". Mistral's own December 2025 human evaluation of Devstral 2 stated that Claude Sonnet 4.5 remained significantly preferred, despite Devstral's competitive open-weight coding performance.
• Current models: Mistral reports strong results for Medium 3.5, Small 4, Leanstral 1.5, OCR 4, and Robostral Navigate on coding, formal verification, document understanding, and robotic-navigation evaluations. These results are task-specific and primarily developer-reported; independent testing is required before applying them to safety-critical or regulated workloads.

Licensing and deployment are separate from benchmark performance. A lower-scoring model may be preferable when local deployment, customization, latency, data residency, or a permissive license is more important than maximum benchmark accuracy.

Application and Ecosystem

Products

• Vibe (formerly Le Chat): Mistral introduced the Le Chat conversational assistant in February 2024. The current official product page states that Le Chat has been renamed Vibe, with existing conversations, settings, and plans carried over. Vibe combines conversational assistance, long-running work agents, coding agents, tool use, search, document processing, and multimodal interaction.
• Mistral AI Studio: The current developer and enterprise environment for building, testing, evaluating, observing, and deploying AI applications, agents, prompts, connectors, and workflows.
• La Plateforme and APIs: La Plateforme was introduced in December 2023 as Mistral's API and customization environment. Its functions have subsequently been incorporated into the broader Studio and API offering. Hosted models can also be accessed through partners including Microsoft Azure, Amazon Web Services, Google Cloud, IBM, and other inference providers, depending on the model.
• Forge: An enterprise system for training, aligning, evaluating, and adapting models to proprietary organizational data and requirements.
• Vibe for Code: Coding-agent interfaces for terminals, integrated development environments, and remotely executed software-engineering tasks. Mistral Vibe CLI was initially introduced with Devstral 2.
• Mistral Compute: Computing infrastructure for training and inference, including Mistral-operated European data-centre capacity. The acquisition of Koyeb and the 2026 data-centre financing formed part of the company's expansion toward a more vertically integrated AI stack.
• Document and Speech Services: Mistral OCR and Document AI process structured and unstructured documents, while the Voxtral family provides transcription, speech understanding, and text-to-speech capabilities.

Enterprise Clients

Mistral AI publishes customer case studies covering financial services, logistics, retail, manufacturing, technology, education, and public-sector organizations. Documented examples include:

• BNP Paribas, which began using Mistral models for Global Markets use cases in the third quarter of 2023 and later expanded the partnership across the group.
• CMA CGM, which applies Mistral systems to shipping, logistics, media, customer-information requests, cargo-release processes, and document processing.
• Zalando, which has used Mistral models through Amazon Bedrock for customer engagement, recommendations, customer service, and operational workflows.
• France Travail, whose agents use language models for summarizing job-seeker records and for an internal retrieval-augmented chatbot covering labour law and benefit policies.

Mistral also lists ASML, Airbus, BMW, HSBC, Stellantis, SAP, Cisco, the European Patent Office, and several governments and public agencies among its customers or strategic partners.

Local, private-cloud, or on-premises deployment can assist organizations with data-residency, confidentiality, and operational-control requirements. It does not automatically ensure compliance with the GDPR or other legislation; compliance also depends on the data, processing purpose, contracts, security controls, retention policy, human oversight, and the organization's legal obligations.

Open-Source Community

Mistral distributes many model weights through Hugging Face and publishes inference, fine-tuning, agent, and integration software through GitHub. Current open-weight models are supported to varying degrees by frameworks including Transformers, vLLM, SGLang, llama.cpp, TensorRT-LLM, and Mistral's own software.

The availability of permissively licensed weights has enabled researchers and developers to create fine-tuned, quantized, multilingual, medical, legal, coding, and domain-specific derivatives. However, the license of the exact base model and every derivative must be checked individually. A model derived from a restricted checkpoint does not become unrestricted merely because its adaptation code is open.

The terms "open-source model" and "open-weight model" are sometimes used interchangeably in company announcements and community discussions. Under the Open Source Initiative's Open Source AI Definition, downloadable weights alone are insufficient: users must also receive the preferred form for modifying the system, including adequate information about training data and the code used to derive the parameters. Mistral generally does not publish the complete pre-training corpora and reproducible end-to-end training pipelines for its frontier models, so "open-weight" is the more precise encyclopedic description.

Licensing

Licensing varies by model and version. The following table gives representative examples rather than a universal rule for every Mistral release.

Model Series or Examples	License	Principal Conditions
Mistral 7B; Mixtral 8x7B and 8x22B; Mistral NeMo; Mathstral; Codestral Mamba; Pixtral 12B; Mistral Small 3; Devstral Small; Magistral Small; Voxtral; Mistral 3; Mistral Small 4; Leanstral	Apache License 2.0	Permits academic and commercial use, modification, and redistribution, subject to preservation of license and attribution notices, patent provisions, and the other terms of Apache 2.0.
Codestral 22B	Mistral AI Non-Production License	Allows research, evaluation, testing, and other non-production use. Production or commercial self-deployment requires a separate commercial agreement.
Mistral Large 2	Mistral Research License	Allows research and non-commercial use and modification. Commercial self-deployment requires a Mistral Commercial License.
Devstral 2 123B; Mistral Medium 3.5	Modified MIT License	Use is not authorized when the user's company, or employer, had global consolidated monthly revenue exceeding US$20 million in the preceding month, unless Mistral grants a separate commercial license. Because of this use restriction, the license is not equivalent to the standard MIT License.
Original Mistral Small, Medium, and Large hosted models; Magistral Medium; OCR and some enterprise models	Proprietary API or negotiated commercial terms	Access, self-hosting, redistribution, and customization depend on the service terms or an enterprise agreement. The existence of an API does not imply that downloadable weights are publicly licensed.

Model family names do not determine licensing. For example, Mistral Large 2 used a research license, whereas Mistral Large 3 was released under Apache 2.0; Devstral Small 2 uses Apache 2.0, whereas the larger Devstral 2 uses the Modified MIT License. The model card and license file for the exact checkpoint should therefore be reviewed before deployment.

A model's lifecycle status is also distinct from its license. Mistral may retire or deprecate a hosted endpoint while previously distributed weights remain usable and redistributable under their original license.

External links

• Official Mistral AI website
• Official Mistral AI documentation
• Current model overview
• Official Mistral AI news and release archive
• Mistral AI on Hugging Face
• Mistral AI on GitHub
• Mistral Vibe, formerly Le Chat
• Mistral AI — Wikipedia

Literature

• Ainslie, Joshua; Lee-Thorp, James; de Jong, Michiel; et al. (2023). GQA: Training Generalized Multi-Query Transformer Models from Multi-Head Checkpoints. arXiv:2305.13245.
• Beltagy, Iz; Peters, Matthew E.; Cohan, Arman (2020). Longformer: The Long-Document Transformer. arXiv:2004.05150.
• Dao, Tri; Fu, Daniel Y.; Ermon, Stefano; Rudra, Atri; Ré, Christopher (2022). FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awareness. arXiv:2205.14135.
• Fedus, William; Zoph, Barret; Shazeer, Noam (2021). Switch Transformers: Scaling to Trillion Parameter Models with Simple and Efficient Sparsity. arXiv:2101.03961.
• Gu, Albert; Dao, Tri (2023). Mamba: Linear-Time Sequence Modeling with Selective State Spaces. arXiv:2312.00752.
• Jiang, Albert Q.; Sablayrolles, Alexandre; Mensch, Arthur; et al. (2023). Mistral 7B. arXiv:2310.06825.
• Jiang, Albert Q.; Sablayrolles, Alexandre; Roux, Antoine; et al. (2024). Mixtral of Experts. arXiv:2401.04088.
• Shazeer, Noam; Mirhoseini, Azalia; Maziarz, Krzysztof; et al. (2017). Outrageously Large Neural Networks: The Sparsely-Gated Mixture-of-Experts Layer. arXiv:1701.06538.