Mistral AI
Last reviewed
May 7, 2026
Sources
62 citations
Review status
Source-backed
Revision
v3 · 8,998 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
May 7, 2026
Sources
62 citations
Review status
Source-backed
Revision
v3 · 8,998 words
Add missing citations, update stale details, or suggest a clearer explanation.
Mistral AI is a French artificial intelligence company headquartered in Paris, founded on April 28, 2023, by Arthur Mensch, Guillaume Lample, and Timothee Lacroix. The three co-founders met during their studies at Ecole Polytechnique and previously worked at Google DeepMind and Meta AI, bringing deep expertise in large language model research to their new venture. Since its founding, Mistral AI has grown into one of Europe's most prominent AI companies, releasing a series of open-weight and proprietary models that compete directly with offerings from OpenAI, Anthropic, and Google.
Mistral AI is notable for its commitment to releasing high-performance models under permissive open-source licenses, its emphasis on European data sovereignty, and its rapid pace of model development. As of May 2026, the company employs roughly 860 people, has a post-money valuation of about EUR 11.7 billion (around $13.8 billion), and is on track to surpass EUR 1 billion in annual recurring revenue by the end of the year [1][2].
| Field | Detail |
|---|---|
| Legal name | Mistral AI SAS |
| Type | Private |
| Industry | Artificial intelligence |
| Founded | April 28, 2023 |
| Founders | Arthur Mensch (CEO), Guillaume Lample (Chief Scientist), Timothee Lacroix (CTO) |
| Headquarters | 15 rue des Halles, Paris, France |
| Other offices | Palo Alto, London, Amsterdam, Marseille |
| Key adviser | Cedric O (former French digital minister) |
| Products | Open-weight and proprietary LLMs, Le Chat (assistant), La Plateforme (API), Mistral Forge (custom models), Mistral Vibe (CLI coding agent) |
| Notable model lines | Mistral, Mixtral, Codestral, Pixtral, Magistral, Devstral, Voxtral, Ministral |
| Funding raised | ~ $3 billion across five rounds (June 2023 to September 2025), plus $830 million in debt (March 2026) |
| Valuation | EUR 11.7 billion ($13.8 billion) post-money, Series C, September 2025 |
| Employees | ~860 (early 2026), targeting 600 additional hires at new Paris HQ |
| Annual recurring revenue | ~ $400 million (January 2026); EUR 1 billion target by end-2026 |
| Website | mistral.ai |
Arthur Mensch, Guillaume Lample, and Timothee Lacroix met during their studies at Ecole Polytechnique and Ecole Normale Superieure, two of France's most prestigious universities [3]. Mensch had spent nearly three years at Google DeepMind, where he focused on transformer architectures and efficient inference methods. Lample and Lacroix had spent several years at Meta AI, contributing to the development of large-scale language models. Lample is recognized as one of the creators of Meta's LLaMA family of models [4].
The trio founded Mistral AI in April 2023, at a time when the global AI industry was experiencing explosive growth following the launch of ChatGPT in late 2022. Arthur Mensch serves as CEO, Guillaume Lample as Chief Scientist, and Timothee Lacroix as CTO. The company is headquartered at 15 rue des Halles in Paris, with additional offices in Palo Alto, London, and Amsterdam, plus an embedded team at CMA CGM's Marseille headquarters [5].
Former French digital minister Cedric O joined Mistral as a non-executive co-founder and adviser shortly after the company was incorporated, a move that has periodically generated debate over potential conflicts of interest, particularly given his earlier role overseeing French tech policy [4].
Mistral AI has raised over $3 billion in equity across five rounds in roughly 29 months, plus an additional $830 million in debt financing in March 2026, making it one of the fastest-growing startups in European history.
| Round | Date | Amount | Valuation | Lead investors | Notes |
|---|---|---|---|---|---|
| Seed | June 2023 | ~ $113M (EUR 105M) | Not disclosed | Lightspeed Venture Partners | Largest seed round in European history at the time; investors included Eric Schmidt, Xavier Niel, JCDecaux Holding [3][6] |
| Series A | December 2023 | ~ $415M (EUR 385M) | ~ $2.1B (EUR 2B) | Andreessen Horowitz, Lightspeed | Brought total funding past $500M; included NVIDIA and Salesforce Ventures [7] |
| Series A extension | February 2024 | ~ $16.3M (EUR 15M) | ~ $2B | Microsoft | Convertible note structured to convert into equity in a subsequent round; tied to the Azure partnership [8] |
| Series B | June 2024 | $640M (EUR 600M) | $6.2B (EUR 5.8B) | General Catalyst, DST Global | Comprised ~ $503M in equity and ~ $137M in debt; investors included NVIDIA, Samsung, Salesforce, IBM, Cisco [9] |
| Series C | September 2025 | ~ $2B (EUR 1.7B) | ~ $13.8B (EUR 11.7B) | ASML | ASML invested EUR 1.3B for an 11% stake; other participants included NVIDIA, DST Global, Andreessen Horowitz, Bpifrance, General Catalyst, Index Ventures, Lightspeed [1][10] |
| Debt facility | March 2026 | $830M | n/a | Bpifrance, BNP Paribas, Credit Agricole CIB, HSBC, La Banque Postale, MUFG, Natixis CIB | Funds purchase of 13,800 NVIDIA GB300 GPUs for a Bruyeres-le-Chatel data center scheduled to come online in Q2 2026 [11] |
The pace at which Mistral accumulated funding was remarkable even by Silicon Valley standards. The company achieved a $2 billion valuation just eight months after founding, and reached a $6 billion valuation within 14 months. The September 2025 Series C more than doubled its previous valuation, driven by a landmark investment from Dutch semiconductor equipment giant ASML. ASML's EUR 1.3 billion contribution made the company one of Mistral's top shareholders with an 11% stake on a fully diluted basis. ASML's CFO Roger Dassen joined Mistral's strategic committee board as part of the deal [1].
While these figures are impressive, they remain considerably smaller than those of American competitors. OpenAI has raised over $57 billion at a $500 billion valuation, and Anthropic has raised approximately $45 billion at a $350 billion valuation [12]. Mistral's strategy has always centered on capital efficiency: building competitive models with smaller teams and lower budgets, then scaling infrastructure once revenue justifies it.
Notable investors across all rounds include Lightspeed Venture Partners, Andreessen Horowitz, NVIDIA, Samsung Venture Investment Corporation, Salesforce Ventures, DST Global, Index Ventures, Bpifrance, IBM, ASML, Eric Schmidt, and Xavier Niel.
From its earliest days, Mistral AI distinguished itself by releasing models with open weights under permissive licenses, most commonly Apache 2.0. This approach allows researchers, developers, and businesses to download, modify, fine-tune, and deploy the models on their own infrastructure without restrictions. The company's first release, Mistral 7B, set the tone: it was distributed via a torrent magnet link on social media, with no formal announcement or press release, and quickly drew attention in the AI community [13].
Mistral has maintained this open-weight philosophy across most of its model family, though some larger and more capable models (such as the original Mistral Large) have been released under more restrictive research licenses or kept entirely proprietary. The general approach is to release smaller, efficient models under Apache 2.0 while monetizing access to frontier models through API services and enterprise contracts.
Mistral AI has positioned itself as a champion of European technological sovereignty in AI. CEO Arthur Mensch has consistently argued that European governments and businesses should not be entirely dependent on American AI providers for critical infrastructure [14]. This message resonates strongly with European regulators, defense agencies, and enterprises that operate under strict data protection regimes such as GDPR.
The company's "private AI" approach, where organizations can host and run Mistral models on their own servers, appeals to heavily regulated industries including finance, healthcare, government, and defense. Mistral has also emphasized multilingual capabilities across European languages, with strong performance in French, German, Spanish, Italian, Portuguese, and other languages [15].
Mensch has stated that he does not believe AI will be dominated by a single winner or country. Instead, he expects multiple regional centers of expertise shaped by local needs, industries, and political realities [14].
Mistral pursues two parallel tracks. Open-weight releases (Mistral 7B, Mixtral 8x7B, Mixtral 8x22B, Mistral Small 3, Mistral Saba, Devstral, Magistral Small, Mistral Large 3, Mistral Small 4, Voxtral, and others) democratize access, drive community adoption, and establish technical credibility. Proprietary, API-only or commercially licensed models (Mistral Large, Mistral Medium 3, Magistral Medium, Pixtral Large, Mistral Medium 3.5) target paying enterprise customers and frontier benchmarks. The mission, repeatedly stated by Mensch, is to "put frontier AI in the hands of everyone" [4].
Mistral AI has maintained an aggressive release cadence since its founding, shipping new models every few months. Between March 16 and March 31, 2026, the company shipped six products in 15 days, averaging one significant release every two and a half days [16]. The following table summarizes the company's major model releases.
| Model | Release date | Parameters | Architecture | License | Key details |
|---|---|---|---|---|---|
| Mistral 7B | September 27, 2023 | 7.3B | Dense transformer with grouped-query attention and sliding window attention | Apache 2.0 | First model; outperformed LLaMA 2 13B on all benchmarks tested; distributed via BitTorrent [13] |
| Mixtral 8x7B | December 11, 2023 | 46.7B total, 12.9B active | Sparse mixture of experts (8 experts, 2 active per token) | Apache 2.0 | Outperformed LLaMA 2 70B on most benchmarks with about 6x faster inference [17] |
| Mistral Large | February 26, 2024 | Not disclosed | Dense transformer | Proprietary | First commercial-only model; launched alongside Le Chat and a partnership with Microsoft Azure [8] |
| Mixtral 8x22B | April 10, 2024 | 141B total, 39B active | Sparse MoE (8 experts, 2 active) | Apache 2.0 | 64K context window; strong on coding and math; released via a torrent link on X [18] |
| Codestral | May 29, 2024 | 22B | Dense transformer | Mistral Non-Production License | First dedicated code model; trained on 80+ programming languages; 32K context window [19] |
| Codestral Mamba | July 2024 | 7B | Mamba2 state-space model | Apache 2.0 | Linear-time inference; designed for very long code contexts [20] |
| Mistral Large 2 | July 24, 2024 | 123B | Dense transformer | Mistral Research License | 128K context, 80+ coding languages, multilingual; scored 84.0% on MMLU [21] |
| Mistral NeMo | July 18, 2024 | 12B | Dense transformer | Apache 2.0 | Built with NVIDIA; 128K context; introduced the Tekken tokenizer [22] |
| Pixtral 12B | September 11, 2024 | 12B + 400M vision encoder | Multimodal (vision + text) | Apache 2.0 | First multimodal model; supports arbitrary image sizes; available on Hugging Face and via torrent [23] |
| Pixtral Large | November 18, 2024 | 124B (123B decoder + 1B vision encoder) | Multimodal | Mistral Research License | Based on Mistral Large 2; 128K context; up to 30 high-res images per input; outperformed GPT-4o on MathVista (69.4%) [24] |
| Codestral 25.01 | January 2025 | Not disclosed | Dense transformer | Proprietary | 256K context; 2x faster than original; #1 on LMSys Copilot Arena [25] |
| Mistral Small 3 | January 30, 2025 | 24B | Dense transformer | Apache 2.0 | Latency-optimized; competitive with LLaMA 3.3 70B at 3x speed; 32K context; 81%+ MMLU [26] |
| Mistral Saba | February 17, 2025 | 24B | Dense transformer | Apache 2.0 | First regional language model; specialized for Arabic and South Indian languages including Tamil [27] |
| Mistral OCR | March 2025 | Not disclosed | Document understanding model | API only | First dedicated OCR model; ~1,000 pages per dollar [28] |
| Mistral Small 3.1 | March 17, 2025 | 24B | Multimodal | Apache 2.0 | Added vision to Small 3; 128K context; 150 tokens/sec inference [29] |
| Mistral Medium 3 | May 7, 2025 | Not disclosed | Dense transformer (multimodal) | Proprietary | Frontier-class at 8x lower cost than competitors; 128K context [30] |
| Devstral | May 21, 2025 | 24B | Dense transformer | Apache 2.0 | Agentic coding model built with All Hands AI; 46.8% on SWE-bench Verified; runs on a single RTX 4090 [31] |
| Magistral Small | June 10, 2025 | 24B | Dense transformer | Apache 2.0 | First reasoning model; chain-of-thought across global languages; 70.7% on AIME 2024 [32] |
| Magistral Medium | June 10, 2025 | Not disclosed | Dense transformer | Proprietary | Enterprise reasoning model; 73.6% on AIME 2024, 90% with majority voting @64 [32] |
| Mistral Small 3.2 | June 20, 2025 | 24B | Multimodal | Apache 2.0 | Maintenance update; better instruction following, fewer repetition errors [33] |
| Voxtral (Small / Mini) | July 16, 2025 | 24B and 3B | Audio + text speech understanding | Apache 2.0 | First Mistral audio model; outperformed Whisper large-v3, GPT-4o mini Transcribe, and Gemini 2.5 Flash on transcription [34] |
| Codestral 2508 | July 2025 | Not disclosed | Dense transformer | Proprietary | 256K context; updated FIM and test generation [35] |
| Mistral Large 3 | December 2, 2025 | 675B total, 41B active | Sparse MoE | Apache 2.0 | Trained from scratch on 3,000 NVIDIA H200 GPUs; first MoE since Mixtral series; image understanding [16] |
| Ministral 3 (3B / 8B / 14B) | December 2, 2025 | 3B, 8B, 14B | Dense transformer | Apache 2.0 | Edge and local deployment; base, instruct, and reasoning variants with image understanding [16] |
| Devstral 2 | December 9, 2025 | 123B (large), 24B (small) | Dense transformer | MIT / Apache 2.0 | Coding-focused; 256K context; 72.2% on SWE-bench Verified [36] |
| Mistral OCR 3 | December 19, 2025 | Not disclosed | Document understanding | API only | 74% win rate over OCR 2; processes up to 2,000 pages per minute on a single GPU; $2 per 1,000 pages [37] |
| Voxtral Transcribe 2 (Mini, Realtime) | February 2026 | 4B class | Speech-to-text | Apache 2.0 | Sub-200ms streaming latency; on-device deployment for cents per hour [38] |
| Leanstral | March 16, 2026 | 120B total, 6B active | Sparse MoE | Apache 2.0 | First open-source Lean 4 proof agent; outperformed larger open-source rivals on FLTEval benchmark [39] |
| Mistral Small 4 | March 16, 2026 | 119B total, 6B active (8B with embeddings) | Sparse MoE (128 experts, 4 active) | Apache 2.0 | Unifies instruct, reasoning, multimodal, and coding; configurable reasoning effort; 256K context [40] |
| Voxtral TTS | March 26, 2026 | 4B | Text-to-speech | CC BY-NC 4.0 (open weights), commercial via API | Voice cloning from under 5 seconds of audio; supports 9 languages; preferred over ElevenLabs Flash v2.5 in 68.4% of blind comparisons [41] |
| Mistral Medium 3.5 | April 30, 2026 | 128B | Dense transformer (multimodal) | Modified MIT (open weights) | Replaces Magistral, Devstral 2, and Medium 3 in a single set of weights; 256K context; 77.6% on SWE-Bench Verified; configurable reasoning_effort [42] |
Released on September 27, 2023, Mistral 7B was the company's debut model and immediately made waves in the AI community. Despite having only 7.3 billion parameters, the model outperformed Meta's LLaMA 2 13B on all benchmarks tested and matched LLaMA 2 34B on several benchmarks [13]. It used grouped-query attention and sliding window attention to achieve efficient inference, with each layer attending to the previous 4,096 hidden states.
The model was released under the Apache 2.0 license and distributed unconventionally via a BitTorrent magnet link posted to social media, without a traditional press release or blog post. The torrent distribution sparked debate about responsible release practices but also signaled the company's commitment to truly open weights, and the approach quickly became a signature for Mistral's early releases [43].
In December 2023, Mistral released Mixtral 8x7B, a sparse mixture-of-experts model. The architecture consists of 8 feedforward expert blocks per layer, with a router network selecting 2 experts per token. While the model has 46.7 billion total parameters, only about 12.9 billion are active during inference for any given token. This design allows the model to match or exceed the performance of much larger dense models while running significantly faster [17]. Mixtral 8x7B outperformed LLaMA 2 70B on most benchmarks with roughly 6x faster inference speed.
Released in April 2024, Mixtral 8x22B scaled the mixture-of-experts approach to 141 billion total parameters with 39 billion active per token. The model offered a 64K token context window and excelled in coding and mathematical tasks. Like its predecessor, it was distributed via a torrent link on X [18].
Mistral Large, released on February 26, 2024, marked the company's first move into proprietary, API-only models. It launched alongside Le Chat (Mistral's consumer chatbot) and a partnership with Microsoft Azure. The model was positioned as competitive with GPT-4, placing Mistral among a small number of companies capable of building frontier-class models. It was fluent in French, German, Spanish, and Italian as well as English [8].
Mistral Large 2, released on July 24, 2024, was a substantial upgrade. The 123-billion-parameter dense model featured a 128K token context window and support for over 80 programming languages. It scored 84.0% on the MMLU benchmark, rivaling all open models except Meta's much larger LLaMA 3 405B. The model showed particular strength in code generation, mathematics, and multilingual reasoning. It was released under the Mistral Research License, allowing use for research and non-commercial purposes [21].
On May 29, 2024, Mistral introduced Codestral, its first model designed specifically for code generation. The 22-billion-parameter model was trained on a diverse dataset covering more than 80 programming languages, including Python, Java, C++, JavaScript, and Bash. With a 32K context window (larger than many competitors at the time), it excelled at code completion, test generation, and fill-in-the-middle tasks. Codestral was integrated into development tools including Continue.dev, Tabnine, LangChain, and LlamaIndex [19].
In July 2024, Mistral followed with Codestral Mamba, a 7-billion-parameter model built on the Mamba2 state-space architecture rather than a transformer. The Mamba design enables linear-time inference and theoretically infinite-length sequence modeling, both of which are valuable for working through long codebases without paying quadratic attention costs [20].
An updated version, Codestral 25.01, arrived in January 2025 with a significantly expanded 256K context window and roughly twice the generation speed of the original. The updated model debuted at the top of the LMSys Copilot Arena leaderboard and was made available through Google Vertex AI, Azure AI Foundry, and GitHub Models [25]. A further refresh, Codestral 2508, landed in mid-2025 and added improved fill-in-the-middle behavior plus stronger test generation in the same 256K context [35].
Released on July 18, 2024, Mistral NeMo was built in collaboration with NVIDIA. This 12-billion-parameter model featured a 128K context window and introduced the Tekken tokenizer, based on Tiktoken and trained on over 100 languages. Tekken was approximately 30% more efficient at compressing text in several European and Asian languages compared to previous tokenizers [22].
The model was trained with quantization awareness, enabling FP8 inference without performance degradation. It was designed as a drop-in replacement for the original Mistral 7B and was released under Apache 2.0.
In September 2024, Mistral released Pixtral 12B, its first multimodal model capable of processing both text and images. Built on the NeMo 12B architecture with a dedicated 400M-parameter vision encoder supporting 1024x1024 image resolution, Pixtral could handle an arbitrary number of images of arbitrary sizes within a single prompt. The model was released under Apache 2.0 and was available on Hugging Face and via torrent [23].
In November 2024, Mistral followed up with Pixtral Large, a 124-billion-parameter multimodal model combining a 123B decoder (based on the Mistral Large 2 architecture) with a 1B parameter vision encoder called Pixtral-ViT. The model could process up to 30 high-resolution images per input within its 128K token context window, equivalent to roughly a 300-page book. On the MathVista benchmark, Pixtral Large achieved 69.4%, outperforming GPT-4o and Gemini 1.5 Pro. It also surpassed those models on ChartQA and DocVQA benchmarks for document and chart understanding. Pixtral Large was released under the Mistral Research License and the Mistral Commercial License [24].
The Mistral Small 3 line went through several iterations in the first half of 2025, evolving from a text-only model to a fully multimodal offering.
Mistral Small 3, released on January 30, 2025, was a 24-billion-parameter dense model that marked a significant licensing shift. While previous Mistral Small models had used the Mistral Research License, Small 3 was released under Apache 2.0, allowing unrestricted commercial use. The model was competitive with Meta's LLaMA 3.3 70B while being over 3x faster on the same hardware, with a 32K context window and over 81% accuracy on MMLU. It offered native function calling and JSON output capabilities [26].
Mistral Small 3.1, released on March 17, 2025, added state-of-the-art vision understanding to the Small 3 foundation. The context window was expanded to 128K tokens, and the model could process images alongside text for tasks like document verification, visual inspection, and image-based customer support. Despite the added multimodal capabilities, the model maintained inference speeds of 150 tokens per second and could run on devices with 32 GB of RAM [29].
Mistral Small 3.2, released on June 20, 2025, was a maintenance update focused on reliability rather than new capabilities. It reduced the rate of infinite or repetitive text generations from 2.11% to 1.29%, improved instruction following accuracy from 82.75% to 84.78% on internal benchmarks, and upgraded the function calling template for better compatibility with frameworks like vLLM [33].
Released on February 17, 2025, Mistral Saba was the company's first specialized regional language model. The 24-billion-parameter model was trained on curated datasets from the Middle East and South Asia, with particular strength in Arabic and South Indian languages such as Tamil. On Arabic-language benchmarks including MMLU, TyDiQA-GoldP, Alghafa, and Hellaswag, Saba outperformed models over five times its size, including LLaMA 3.1 70B and Jais 70B [27].
The model was lightweight enough to deploy on single-GPU systems, responding at speeds exceeding 150 tokens per second. It was released under Apache 2.0 and targeted use cases including Arabic-language conversational AI, fine-tuning for industries such as energy and healthcare, and localized content generation. Saba represented Mistral's strategy of building regional expertise rather than treating all languages as secondary to English.
Released on May 7, 2025, Mistral Medium 3 was a frontier-class dense language model designed for enterprise workloads. Mistral positioned it as delivering performance at or above 90% of Claude Sonnet 3.7 on benchmarks across the board, while costing up to 8x less (at $0.40 per million input tokens and $2 per million output tokens). The model supported a 128K token context window and both text and image inputs, with particular strength in coding, STEM reasoning, and multimodal understanding [30].
Mistral Medium 3 could be deployed on any cloud environment, including self-hosted setups with as few as four GPUs. It represented the company's push into the enterprise market with a model that balanced frontier-level capability against practical cost and deployment constraints.
The successor, Mistral Medium 3.5, arrived on April 30, 2026, and represented a major consolidation of Mistral's product line. It is a dense 128-billion-parameter model with a 256K context window, released under a modified MIT license with open weights on Hugging Face. The release retired Magistral (reasoning), Devstral 2 (coding), and Medium 3 (chat), folding all three product lines into one set of weights with a configurable reasoning_effort parameter. The model scored 77.6% on SWE-Bench Verified, runs on four GPUs, and is priced at $1.50 per million input tokens and $7.50 per million output tokens through Mistral's API. The vision encoder was retrained from scratch to handle variable image sizes and aspect ratios. Medium 3.5 became the default model in both Le Chat and Mistral Vibe, and powered the launch of remote agents in the Vibe coding platform [42][43].
Mistral released two generations of its agentic coding model line in 2025.
Devstral, released on May 21, 2025, was built through a collaboration between Mistral AI and All Hands AI. The 24-billion-parameter model was designed specifically for real-world software engineering tasks rather than isolated code generation. It achieved 46.8% on SWE-bench Verified, outperforming prior open-source state-of-the-art models by more than 6 percentage points. Unlike typical code completion models, Devstral was trained to solve actual GitHub issues by contextualizing code within large codebases and identifying relationships between components. It ran on agent scaffolds such as OpenHands and SWE-Agent. The model was small enough to run on a single RTX 4090 or a Mac with 32 GB RAM, and was released under Apache 2.0 [31].
Devstral 2, released on December 9, 2025, was available in two sizes. Devstral 2 (123B parameters) was a dense transformer with a 256K context window that scored 72.2% on SWE-bench Verified, placing it among the best open-weight coding models. Devstral Small 2 (24B parameters) scored 68.0% on SWE-bench Verified while being small enough to run on consumer hardware [36].
Alongside Devstral 2, Mistral introduced Mistral Vibe, a native command-line interface designed for end-to-end code automation. The CLI is capable of understanding entire codebases rather than just individual files, enabling architecture-level reasoning [36]. In May 2026, Vibe gained remote agents that run in the cloud rather than on the user's laptop. Sessions can be teleported from local to remote with task state preserved, run in isolated sandboxes, and open pull requests on GitHub when complete. Both Vibe and Le Chat default to Mistral Medium 3.5 for these agentic flows [44].
On June 10, 2025, Mistral released Magistral, its first family of reasoning models. The Magistral line was designed to compete with OpenAI's o1 and o3 series and DeepSeek's R1 by performing explicit chain-of-thought reasoning before arriving at answers.
Two variants were released. Magistral Small was a 24B parameter open-source model available under Apache 2.0, and Magistral Medium was a more powerful enterprise version available through the API. Magistral Medium scored 73.6% on AIME 2024 (a challenging math competition benchmark), reaching 90% accuracy with majority voting at 64 samples. Magistral Small scored 70.7% on the same benchmark [32].
A distinguishing feature of the Magistral family was its multilingual chain-of-thought capability. The models could reason across global languages and alphabets, including English, French, Spanish, German, Italian, Arabic, Russian, and Simplified Chinese. In Le Chat, the Magistral models powered a new "Think" mode for deep reasoning and "Flash Answers" for responses at up to 10x the speed of competitors [32]. Magistral was retired in April 2026 when its reasoning capability was absorbed into Mistral Medium 3.5.
In July 2025, Mistral released Voxtral, its first audio understanding model. The family includes a 24B variant for production-scale applications and a 3B variant for local and edge deployments, both under Apache 2.0. Voxtral handles audios up to 30 minutes for transcription and 40 minutes for understanding within a 32K token context, supports built-in question answering and summarization over audio, and is natively multilingual. On benchmarks, Voxtral outperformed OpenAI's Whisper large-v3, GPT-4o mini Transcribe, and Gemini 2.5 Flash on transcription tasks [34].
In February 2026, Mistral released Voxtral Transcribe 2 in two flavors: Voxtral Mini Transcribe V2 for batch transcription, and Voxtral Realtime for live applications with latency configurable down to sub-200ms. Both run on-device for what Mistral describes as cents per hour of audio [38].
Mistral completed its voice stack in March 2026 with Voxtral TTS, a 4-billion-parameter text-to-speech model that runs on a single GPU with 16 GB VRAM. The model can clone voices from under 5 seconds of reference audio, supports 9 languages (English, French, German, Spanish, Dutch, Portuguese, Italian, Hindi, Arabic), and was preferred over ElevenLabs Flash v2.5 in 68.4% of blind comparisons in human listening tests, with the gap widest in Spanish (87.8%) and Hindi (around 80%). Open weights are released under CC BY-NC 4.0 for research and non-commercial use, with commercial access via the Mistral API at $0.016 per 1,000 characters of generated audio [41].
In March 2025, Mistral introduced Mistral OCR, a dedicated document understanding API priced at roughly 1,000 pages per dollar. The original model targeted enterprise document workflows that combine text extraction with structured layout understanding [28].
The successor, Mistral OCR 3 (mistral-ocr-2512), launched on December 19, 2025. It achieved a 74% overall win rate over Mistral OCR 2 on forms, scanned documents, complex tables, and handwriting. The model handles cursive and mixed-content annotations, is robust to compression artifacts and skew, and reconstructs table structures with merged cells using HTML colspan and rowspan tags. OCR 3 processes up to 2,000 pages per minute on a single GPU and is priced at $2 per 1,000 pages, with a 50% discount for batch processing. Access is through the API or the new Document AI Playground in Mistral AI Studio [37].
On December 2, 2025, Mistral AI released the Mistral 3 model family, representing a major step forward in capabilities.
Mistral Large 3 is a sparse mixture-of-experts model with 675 billion total parameters and 41 billion active parameters per token. It was trained from scratch on 3,000 NVIDIA H200 GPUs. This was Mistral's first mixture-of-experts frontier model since the original Mixtral series, and the company described it as a substantial step forward in pretraining. The model achieves competitive performance with the best instruction-tuned open-weight models on general tasks, while also adding image understanding and best-in-class multilingual performance. It debuted at number two in the open-source non-reasoning models category on the LMArena leaderboard. It was released under the Apache 2.0 license [16].
Ministral 3 is a family of smaller models designed for edge computing and local deployment, available in three sizes: 3B, 8B, and 14B parameters. For each size, Mistral released base, instruct, and reasoning variants, all with image understanding capabilities and all under Apache 2.0. The company specifically targeted use cases in drones, cars, robots, phones, and laptops, aiming for the best cost-to-performance ratio among open-source small models [16].
Released on March 16, 2026, Leanstral is the first open-source AI agent built specifically for Lean 4 formal verification. The 120-billion-parameter model operates on just 6 billion active parameters and is designed to generate and verify mathematical proofs, a capability with applications in securing smart contracts, blockchain protocols, and safety-critical software [39].
Mistral introduced a new benchmark called FLTEval to evaluate formal proof engineering, and reported that Leanstral outperformed larger open-source models including GLM5-744B-A40B and Qwen3.5-397B-A17B on this benchmark. Leanstral is available with open weights under Apache 2.0 as an agent mode within Mistral Vibe and through a free API endpoint.
Released on March 16, 2026, Mistral Small 4 is a 119-billion-parameter mixture-of-experts model with 128 experts and 4 active experts per token. With only 6 billion active parameters per token (8 billion including embedding and output layers), it is remarkably efficient for its capability level [40].
The model unifies four previously separate capabilities into a single deployment: instruction following (previously Mistral Small), reasoning (previously Magistral), multimodal understanding (previously Pixtral), and agentic coding (previously Devstral). It supports a 256K context window and introduces a configurable reasoning_effort parameter that allows developers to trade latency for deeper test-time reasoning on a per-request basis.
Mistral Small 4 delivers a 40% reduction in end-to-end completion time and 3x more requests per second in throughput-optimized configurations compared to Mistral Small 3. It is released under the Apache 2.0 license.
Le Chat is Mistral AI's consumer-facing chatbot and AI assistant, comparable to ChatGPT or Claude. The name (pronounced /lə ʃa/ in French) means "the cat." It was first launched alongside Mistral Large in February 2024 and has undergone significant updates since [8].
In November 2024, Mistral overhauled Le Chat with a new interface that included web search with inline citations, document and image analysis (supporting PDFs with graphs and equations), image generation powered by Black Forest Labs' Flux Pro model, and a Canvas interface for collaborative editing of documents, presentations, code, and mockups. Canvas allows users to modify content in place without regenerating entire responses, version drafts, and preview designs. Task agents were also introduced, enabling complex multi-step workflows [45].
On February 6, 2025, Mistral released Le Chat on iOS and Android, with conversation synchronization across web and mobile platforms. Alongside the mobile launch, the company introduced a Pro subscription tier at $14.99 per month, positioned as the most affordable premium AI subscription compared to ChatGPT Plus at $20 and Claude Pro at $20. Le Chat Pro provides unlimited access to Mistral's frontier models, uncapped daily messages, priority model access, enhanced image generation, and a "No Telemetry Mode" that prevents prompts from being used for model training [46].
As of mid-2026, Le Chat is offered in four tiers:
| Tier | Price | Key features |
|---|---|---|
| Free | $0 | Latest models, ~25 messages per day cap, basic web search |
| Pro | $14.99 / user / month | Unlimited chat (soft cap ~150 messages/day), all frontier models, 150 Flash Answers per day, No Telemetry Mode, up to 500 saved memories, 15 GB document storage |
| Team | $24.99 / user / month | Pro features plus unified billing, admin controls, priority support, up to 30 GB storage per user |
| Enterprise | Custom (contracts typically start ~$20,000 / month) | Self-hosted or private-cloud deployment, SAML SSO, audit logs, data export, admin API, custom models, dedicated SLAs |
Le Chat also includes a memory system that retains context across conversations and can learn user preferences over time. Users can create custom Agents within Le Chat, granting them permissions for web search, canvas drawing, image generation, code execution, and integrations with Gmail and Google Calendar. The Agents API supports building enterprise-grade agentic AI systems, and early testers reported 25-35% lower first-token latency compared to OpenAI's GPT Builder or Google's Gemini Gems [45].
Le Chat Enterprise launched as a separate, fully customizable tier in mid-2025 and is the focus of much of Mistral's enterprise revenue. Mistral reported tripling its revenue within 100 days of launching the enterprise version, with rapid uptake driven by private cloud hosting, AI agent builders for custom workflows, and integrations with internal document systems via the Model Context Protocol (MCP) [47]. In May 2026, Le Chat added a "Work mode" that uses remote agents and Mistral Medium 3.5 to handle long, multi-step tasks while users continue with other work [44].
La Plateforme is Mistral's API service for developers and businesses. It provides access to the full Mistral model range through chat completion endpoints, embedding endpoints, and specialized APIs for code generation, OCR, audio, and agent orchestration [48].
The platform offers models at different performance and price tiers, from smaller, cheaper models suitable for high-throughput applications to frontier models for complex reasoning. La Plateforme supports pay-per-token pricing, fine-tuning, multiple deployment options (cloud, on-premises, edge), and both instruction-following and fill-in-the-middle code generation. It is available directly from Mistral and through major cloud providers including Amazon Bedrock, Microsoft Azure AI, Google Vertex AI, GitHub Models, Amazon SageMaker, and IBM watsonx.
Announced at NVIDIA GTC on March 17, 2026, Mistral Forge is an enterprise platform that allows organizations to build custom AI models trained on their own proprietary data. The platform supports three stages of model customization: pre-training for building domain-aware models, post-training for refining models for specific tasks, and reinforcement learning for improving agentic performance in real-world environments [2].
Forge enables enterprises to train models on internal documentation, codebases, structured data, and operational records. Early customers include Ericsson, the European Space Agency, Italian consulting firm Reply, and Singapore's DSO and HTX defense organizations [2].
CEO Mensch has described Forge as central to Mistral's enterprise strategy, positioning the company as infrastructure for organizations that want to own their AI rather than rent it from cloud providers.
Mistral Vibe is a native CLI coding agent introduced alongside Devstral 2 in December 2025. It interprets entire codebases rather than individual files, runs in isolated sandboxes, and integrates with GitHub for pull request creation. The May 2026 update added remote agents that run in the cloud, with the ability to teleport ongoing local sessions to remote instances and continue them in parallel [44]. Vibe defaults to Mistral Medium 3.5 from May 2026 onward.
Mistral's Document AI stack combines Mistral OCR 3 for layout-aware extraction with the Document AI Playground in Mistral AI Studio. The stack targets enterprise contract review, scanned-form digitization, and large-volume archival projects, with the headline pricing of $2 per 1,000 pages making it competitive against legacy OCR vendors [37].
Mistral AI operates in a fiercely competitive market dominated by well-funded American companies. OpenAI, the market leader, generated approximately $11.9 billion in revenue in 2025 and is targeting $30 billion in 2026. Anthropic reached roughly $7 billion in annualized revenue by late 2025. By comparison, Mistral reported approximately EUR 300 million in annual recurring revenue as of September 2025, then accelerated sharply to roughly $312 million ARR by December 2025 and $400 million ARR by January 2026, with the company projecting EUR 1 billion (approximately $1.2 billion) by the end of 2026 [12][49].
The funding gap is equally stark. OpenAI has raised over $57 billion, and Anthropic over $45 billion, while Mistral's total equity funding stands at roughly $3 billion plus an additional $830 million in debt for compute infrastructure [12]. Despite these disparities, Mistral has carved out a viable position by focusing on several areas of differentiation.
First, open-weight models with permissive licenses attract developers and enterprises that want to avoid vendor lock-in. Second, the European identity appeals to governments and regulated industries that prefer non-American AI providers, with about 60% of Mistral's revenue coming from Europe [49]. Third, the company's focus on efficient, cost-effective models, particularly through mixture-of-experts architectures and the new dense Medium 3.5, allows it to serve enterprise use cases where running a smaller, fine-tuned model locally is more practical than paying for API access to a massive frontier model.
Mistral also faces growing competition from Chinese AI labs, particularly DeepSeek, whose R1 reasoning model achieved strong benchmark results at very low cost in early 2025. The emergence of capable, low-cost Chinese models has intensified the pressure on all Western AI companies to deliver better performance per dollar. Mistral competes with Meta AI's open-source LLaMA family in the open-weight space and with OpenAI, Anthropic, and Google in the commercial API and enterprise markets, including against newer frontier models like GPT-5 and Claude Opus 4.7.
Mistral has been a leading proponent of mixture-of-experts (MoE) architectures in production language models. The Mixtral series demonstrated that sparse MoE models could match or exceed the performance of much larger dense models at a fraction of the computational cost. The approach has since been adopted or explored by other major AI labs. Mistral Large 3 (675B total, 41B active) and Mistral Small 4 (119B total, 6B active) represent the company's most advanced MoE deployments to date [16][40].
Mistral's original 7B model popularized the use of sliding window attention (SWA) at scale, with each layer attending to the previous 4,096 hidden states. This pattern reduces memory usage and supports efficient processing of longer sequences without the quadratic cost of dense attention [13].
Codestral Mamba was one of the first production code models to use the Mamba2 state-space architecture. The design offers linear-time inference and the ability to model sequences of effectively unbounded length, which is particularly useful for working through long codebases and continuous streams of tokens [20].
The Tekken tokenizer, introduced with Mistral NeMo, was trained on over 100 languages and demonstrated approximately 30% better compression efficiency for several European and Asian languages compared to standard tokenizers. This improvement directly translates to lower inference costs and longer effective context windows for multilingual applications [22].
Mistral Small 4 (March 2026) and Mistral Medium 3.5 (April 2026) introduced a reasoning_effort parameter that lets developers control the depth of chain-of-thought reasoning on a per-request basis. Some queries benefit from extended deliberation, others need fast responses, and the toggle lets a single deployed model handle both without switching weights [40][42].
With Leanstral (March 2026), Mistral became one of the first major AI labs to release an open-source agent specifically designed for formal mathematical proof verification. By targeting the Lean 4 proof assistant ecosystem, the company opened a path toward AI-assisted verification of software correctness, a domain with growing importance in safety-critical applications and blockchain security [39].
Mistral AI has established partnerships with several major technology companies and enterprises:
| Partner | Date | Details |
|---|---|---|
| Microsoft | February 2024 | Mistral Large available on Azure AI as a first-party offering; Microsoft invested EUR 15M (~ $16M) structured as a convertible note; partnership covers supercomputing, model distribution, and joint R&D [8] |
| NVIDIA | July 2024 onward | Co-developed Mistral NeMo; Mistral Large 3 trained on NVIDIA H200 GPUs; Forge announced at NVIDIA GTC 2026; new Bruyeres-le-Chatel data center uses 13,800 NVIDIA GB300 GPUs [22][2] |
| ASML | September 2025 | EUR 1.3B investment for 11% stake; strategic partnership for European AI sovereignty and semiconductor design AI [1] |
| CMA CGM | April 2025 | EUR 100M five-year partnership; six Mistral employees embedded at CMA CGM's Marseille HQ in a dedicated "Mistral AI Factory"; processes 1 million emails weekly [50] |
| BNP Paribas | 2024, extended February 2026 | Three-year partnership covering all current and future Mistral commercial models; deployed across global markets, sales, and customer support; GDPR-compliant on-premises deployment [51] |
| AXA | 2024 | Empowers 140,000+ employees with secure AI for text generation and analysis using Mistral models [49] |
| Stellantis | October 2025 | Company-wide deployment of a custom industrial Mistral LLM; joint innovation lab and transformation academy [52] |
| Cisco | 2025 | AI agents for customer experience [53] |
| TotalEnergies | 2025 | AI innovation in the energy sector [54] |
| ALTEN | 2025 | AI solutions in engineering services [55] |
| Capgemini | 2024 onward | Generative AI model adoption across consulting engagements [56] |
| Accenture | 2026 | Strategic alliance to accelerate enterprise AI reinvention with sovereign AI capabilities [57] |
| Agence France-Presse | 2025 | Le Chat granted access to AFP's text archives dating back to 1983 [4] |
| Amazon Web Services | Ongoing | Mistral models available on Amazon Bedrock and SageMaker |
| Google Cloud | Ongoing | Models available on Google Vertex AI |
| IBM | 2024 onward; expanded May 2026 | Models available on IBM watsonx; first certified provider in IBM's Sovereign Core ecosystem at Think 2026 [58] |
Other notable enterprise customers include Orange and Ericsson. The Microsoft partnership drew scrutiny from some in the open-source community, who questioned whether a company that had positioned itself as a European open-source alternative should be so closely aligned with a major American tech firm. Mistral has maintained that partnerships with cloud providers are necessary for distribution and do not compromise its independence or open-source commitments. The UK's Competition and Markets Authority cleared the Microsoft partnership in May 2024, finding that Microsoft did not have "material influence" over Mistral [59].
In March 2026, Mistral raised $830 million in debt financing to fund a new data center in Bruyeres-le-Chatel, near Paris. The facility will house 13,800 NVIDIA GB300 GPUs as part of Mistral's Grace Blackwell deployment, with operations expected to begin in Q2 2026. Seven banks backed the transaction: Bpifrance, BNP Paribas, Credit Agricole CIB, HSBC, La Banque Postale, MUFG, and Natixis Corporate & Investment Banking. Mistral has stated plans to scale up to 200 MW of capacity across Europe by the end of 2027 [11].
This represents a meaningful shift in strategy. For most of its history, Mistral relied on rented compute from Microsoft Azure, AWS, and Google Cloud. Owning its own training and inference capacity gives the company more control over costs, latency for European users, and data sovereignty for regulated customers, all of which are central to the sales pitch for Le Chat Enterprise and Mistral Forge.
As a prominent European AI company, Mistral AI has been actively involved in policy discussions around AI regulation, particularly the European Union's AI Act. The company has generally advocated for regulation that focuses on specific applications and use cases rather than regulating foundation models themselves. Mensch has argued that overly restrictive regulation of open-source models could hamper European competitiveness and push AI development further toward the United States and China [14].
Mistral's advocacy for open-weight models aligns with broader arguments in the AI policy community that open access to model weights promotes transparency, enables independent safety research, and prevents excessive concentration of power in a small number of companies. The company has also engaged with industry bodies including the Frontier Model Forum and various European policy initiatives.
Mistral's February 2024 investment and cloud partnership with Microsoft drew attention from the European Commission as part of broader reviews of Big Tech's AI alliances. Both companies stated they would cooperate with information requests, and the UK's Competition and Markets Authority cleared the partnership in May 2024 [59].
Mistral has weathered several public controversies during its rapid expansion.
The Microsoft partnership, announced in February 2024, raised concerns in parts of the open-source community and EU policy circles about reinforcing US tech dominance and deviating from the company's stated open-source principles, particularly because the launch involved a closed-source flagship model [59]. Mistral has continued to release open weights for most subsequent models, but the criticism has not entirely faded.
Former French digital minister Cedric O's role as an adviser and non-executive co-founder has also drawn periodic criticism. Some commentators have questioned whether his earlier government role created conflicts of interest given his current advocacy on AI policy [4].
In early 2025, a French digital rights complaint to CNIL accused Mistral of exploiting users' personal data without proper consent in the free version of Le Chat, citing the absence of clear opt-out controls. Mistral subsequently introduced No Telemetry Mode for Pro subscribers and clarified data handling defaults [60].
In August 2025, an OECD AI Incidents tracker entry recorded allegations from a former Mistral employee that one of the company's recent reasoning models had been distilled from DeepSeek output and that benchmark results had been misrepresented. Mistral has not publicly confirmed the allegations, and no formal action has resulted at the time of writing [61].
Guillaume Lample's earlier work on LLaMA at Meta has been referenced in copyright litigation against Meta, where plaintiffs allege that Meta used pirated books from sites such as LibGen for training. The cases concern Meta's models rather than Mistral's, but have surfaced Lample's name in coverage of the AI training data debate [62]. The initial torrent distribution of Mistral 7B in 2023 was also criticized as unconventional, though it has since been broadly seen as part of Mistral's ethos rather than a controversy [43].
As of May 2026, Mistral AI is in a strong position. The company employs over 860 people, is planning a new Paris headquarters with capacity for 600 additional hires, and has the infrastructure cornerstone of a 13,800-GPU data center coming online in Bruyeres-le-Chatel during Q2 [5][11]. Annual recurring revenue passed $400 million in January 2026 and the company is on pace to clear EUR 1 billion ARR by year-end [49].
The April 2026 release of Mistral Medium 3.5, which collapsed the Magistral, Devstral 2, and Medium 3 product lines into a single dense model, simplified the lineup and made the company easier to sell to enterprises that previously had to choose between specialist offerings. The May 2026 launch of remote agents in Vibe and the Work mode in Le Chat extended the same model into asynchronous, long-running coding workflows. Mistral Forge, announced at NVIDIA GTC in March 2026, positions Mistral as more than a model provider: a full-stack AI infrastructure company for enterprises that want to own and customize their AI systems [42][2].
The company faces significant challenges. American competitors continue to outspend it by an order of magnitude, top AI talent remains expensive globally, efficient Chinese models such as DeepSeek and Qwen put steady pressure on margins, and balancing open-source ideals with commercial revenue generation remains a constant negotiation. Mistral has consistently defied expectations since its founding, growing from three co-founders to a multi-billion-dollar company in under three years, but the next phase, building both its own compute and its own enterprise distribution, is a different and harder kind of bet.