Mistral AI Explained: Large 4, Vibe, and the Open-Model Strategy

Human Verified

Share to: , and

A wide illustration of Paris rooftops at dusk with a flowing wind ribbon and glowing neural-network nodes above the city, a European AI lab. No text, no logos.

Mistral AI Explained: Large 4, Vibe, and the Open-Model Strategy

Mistral AI is the French lab that made openness its business model. Founded in Paris in April 2023, it builds frontier AI models and then does something its American rivals mostly do not: it publishes the weights, so anyone can download the model, inspect it, fine-tune it and run it on their own machines. Its current flagship is Mistral Large 4, nicknamed “Le Chonk” — a trillion-parameter, natively multimodal model whose weights were released openly in late October 2026. Around that flagship sits a full stack: a smaller hybrid model called Mistral Small 4, an AI agent called Vibe that handles both office work and coding, an enterprise platform called Mistral Studio, and specialist models for documents and voice. Checked 21 November 2026 against Mistral’s own documentation and announcements, read in full.

Two things confuse almost everyone who meets Mistral for the first time, and this guide clears both up before anything else. The first is the naming. Mistral’s model names sound like a family — Mistral, Magistral, Codestral, Devstral, Ministral, Pixtral, Voxtral — but they are not a stable family tree. They are a fast-moving catalogue in which whole lines appear, merge and retire within months. Magistral, the reasoning line launched in June 2025, is already retired from the API. Devstral 2, the coding model launched in December 2025, lasted about eight months. The second confusion is what “open” means. Mistral uses the word in two different senses: some models are open source under the Apache 2.0 licence, with almost no restrictions, while the flagship Large 4 is open-weight under Mistral’s own licence, which is free to use and commercialise but requires a separate commercial licence if you sell it as a hosted service. This guide says exactly which models are current, which are retired, and which licence each one carries — because a guide that lists Magistral and Codestral as today’s pillars is already out of date.

There is also a correction that belongs at the top, because the loudest stories about Mistral in 2026 are about money and sovereignty rather than models, and they are easy to misread. In September 2026 Mistral raised €3 billion at a valuation above €21 billion, led by Samsung — the largest equity round ever raised by a European technology company, by the company’s own account. That money is funding its own datacentres in Europe and a pitch it calls sovereign AI: models trained in Europe, hosted in Europe, and runnable on a customer’s own infrastructure. But sovereignty is a strategy, not a guarantee, and this guide treats it that way. It also treats Mistral’s benchmark claims the way they deserve to be treated: they are Mistral’s own numbers, from its own announcements and evaluations it commissioned, and they are labelled as such throughout.

A wide illustration of Paris rooftops at dusk with a flowing wind ribbon and glowing neural-network nodes above the city, a European AI lab. No text, no logos.

What Mistral AI Is — Europe’s Open-Model Lab

<img src="

A bright modern research lab in Paris, engineers at desks with screens showing abstract charts, warm light. No readable words.

” alt=”A bright modern research lab in Paris, engineers at desks with screens showing abstract charts, warm light. No readable words.”>

Mistral AI is a French artificial intelligence company headquartered in Paris. It was founded in April 2023 by three researchers: Arthur Mensch, formerly of Google DeepMind, who is the chief executive; Guillaume Lample, formerly of Meta, who is the chief science officer; and Timothée Lacroix, also formerly of Meta, who is the chief technology officer. The three had met years earlier at École Polytechnique, the French engineering school outside Paris, and two of them had helped build the original Llama models at Meta. That pedigree matters, because Mistral was founded on a specific bet: that the open approach that had already won in web browsers, operating systems and cloud orchestration would win in generative AI too, and that a European lab could build frontier models without copying the closed, API-only playbook of the American labs.

The company’s founding statement, published on 27 September 2023 alongside its first model, is unusually blunt for a launch post, and it still describes the strategy three years later. Mistral argued that proprietary models had revealed the power of generative AI, but that open solutions would “quickly outperform proprietary solutions for most use cases”, driven by community development and the demand for technical excellence. Its mission, in its own words, is “to spearhead the revolution of open models”. The argument was not only philosophical. Open models, Mistral said, let developers control the engine that powers their application, adapt model size and cost to the difficulty of the task, deploy on their own infrastructure, preserve data privacy, and customise guardrails and editorial tone instead of depending on the choices of a black-box provider. Closed APIs, it warned, introduce intellectual-property leakage risks and a “cultural liability”, because the generated content stays under the provider’s control.

The first model, Mistral 7B, arrived the same day as that statement — a seven-billion-parameter model released under the Apache 2.0 licence, which Mistral claimed outperformed all then-available open models up to 13 billion parameters on standard English and code benchmarks. It was a small model by the standards of the time, and that was the point. Mistral’s early reputation was built on efficiency: small models that punched above their size, released openly, with the weights free to download. The Mixtral series followed, then larger models, and the company grew fast. Its own timeline records a seed round in June 2023, a Series A in December 2023, the first Mistral Large in February 2024, a Series B in June 2024, and its hundredth employee in September 2024.

The money got much bigger in 2025 and 2026. In September 2025 Mistral closed a €1.7 billion Series C led by ASML, the Dutch chip-equipment maker, at a valuation of €11.7 billion. A year later, on 8 September 2026, it announced a €3 billion Series D at a post-money valuation of more than €21 billion, led by Samsung Electronics, with the Scaleup Europe Fund managed by EQT and existing investor PSG Equity as co-leads. The company calls it the largest equity fundraising round ever completed by a European technology company, and the investor list reads like a map of its strategy: Advent, funds managed by BlackRock and the Grand Duchy of Luxembourg joined as new investors, while a16z, ASML, Nvidia, Salesforce Ventures and others returned. The Series D announcement says the company now operates in 20 countries and supports more than 125 global enterprises, naming Airbus, ASML and HSBC. Deutsche Telekom’s technology fund has also invested in the company, backing a sovereign-AI push in Germany that this site covered in November 2026.

That funding is being spent on the thing that makes Mistral different from a pure model lab: the full stack. Mistral now builds the models, the infrastructure they run on, and the products that bring them into production. It has its own datacentres in Europe — the Large 4 model was trained on 3,800 Nvidia Grace Blackwell GPUs in Mistral’s own European datacentres — and it sells compute, custom model training, an enterprise platform, and an assistant, alongside the API. The pitch, repeated across its announcements, is that customers should never be locked into a single vendor’s roadmap, pricing or availability. Whether that pitch survives contact with the realities of running a trillion-parameter model is the question this guide keeps returning to, but the shape of the company is clear: a Paris lab that wants to be Europe’s answer to OpenAI, Google and Anthropic, built on openness rather than on a walled garden.

The Model Family at a Glance

A branching tree of glowing nodes and connecting lines on a dark background, a model family tree. No readable labels.

Mistral’s catalogue is best understood as three layers that change at different speeds. At the top is the flagship line, currently Mistral Large 4, which changes roughly once a year and carries the company’s reputation. In the middle is the workhorse line, currently Mistral Small 4, which is where most real usage happens. And around them is a ring of specialist lines — reasoning, coding, documents, voice, embeddings — that appear, merge into the general models, or retire, sometimes within months. The table below is the state of the family on 21 November 2026, taken from Mistral’s own documentation, which lists both the current models and a long table of deprecated and retired ones. The retired column is as informative as the current one: it shows how fast this catalogue turns over.

The current general-purpose models are Mistral Large 4 and Mistral Small 4. Large 4 is the flagship: a trillion-parameter mixture-of-experts model, natively multimodal, with a one-million-token context window, released openly under Mistral’s own open-weight licence. Small 4 is the hybrid workhorse: 119 billion parameters in total with only about 6.5 billion active per token, a 256,000-token context window, released under the permissive Apache 2.0 licence. Between them, until 31 October 2026, sat Mistral Medium 3.5, a frontier-class multimodal model for agentic and coding work. It is now deprecated — Mistral’s documentation says plainly, “Use Mistral Large 4 for new integrations” — and it is scheduled to leave the API entirely at the end of April 2027. Anyone whose code still calls Mistral Medium has a migration to plan.

The specialist lines tell the same story of consolidation. Magistral was Mistral’s first reasoning model, launched in June 2025 in two sizes: a 24-billion-parameter open version and a larger enterprise version. Both are now on the deprecated list, and the reasoning capability they introduced has been absorbed into Small 4 and Large 4 as a setting rather than a separate model. Codestral, the code-completion model, has had two versions: 25.01, retired in November 2025, and 25.08, which is current. Devstral, the agentic coding model, went through three generations in about eighteen months — Devstral Small and Medium, then Devstral 2 and Devstral Small 2 in December 2025 — and Devstral 2 was retired from the API in July 2026. Pixtral, the vision line, has been folded into the general models too: Large 4 has its own vision encoder, and Small 4 accepts image inputs natively, so there is no longer a separate Pixtral to choose.

What remains as genuinely separate specialists are the models that do jobs the general models do not. Mistral OCR 4 handles document understanding — turning PDFs and scans into structured text with bounding boxes and confidence scores. Voxtral 3 handles speech: transcription, text-to-speech and full voice agents. Codestral Embed handles code search. Mistral Embed handles text retrieval. Mistral Moderation handles safety filtering. And Mistral has begun adding models that are not language models at all: Timestral, a small 209-million-parameter model for time-series forecasting, and research lines with names like Leanstral, Robostral and Shieldstral that have so far been announced only in headlines. The pattern is consistent: Mistral launches a specialist, proves the capability, and then either folds it into the flagship or keeps it as a standalone tool for the customers who need exactly that job.

Two other facts belong in any map of the family. First, Mistral’s API now sells other companies’ models alongside its own. Its model list includes Z.ai’s GLM 5.3 and GLM 5.3 Flash, and DeepSeek’s V4.1 Flash, each with a million-token context window. That is a significant shift for a company whose identity is building its own models: the platform is becoming a marketplace as well as a lab. Second, the naming is genuinely confusing even for Mistral’s own customers, because version numbers are dates — 25.08 means August 2025, 26.10 means October 2026 — and because the same name can mean different things in different years. Codestral 25.01 and Codestral 25.08 are different models. Mistral Large 3 and Mistral Large 4 are different models with different licences. The only reliable way to know what you are using is the model identifier in the API, and that is a habit this guide recommends throughout: pin the exact model ID in your configuration, and check the documentation’s deprecated list before you build anything new.

Mistral Large 4 “Le Chonk”: The Flagship

A large, friendly, chunky cat-like robot made of glowing cubes standing on a datacentre floor, the flagship model. No text, no logos.

Mistral Large 4 is the model everything else in Mistral’s current catalogue is measured against. It was announced on 6 October 2026 as a public preview, its weights were released on 29 October 2026, and it reached general availability on 5 November 2026 as version 26.10. Its unofficial name is “Le Chonk”, and Mistral uses the nickname itself — the weights-release announcement is titled “Le Chonk: Built for Coders, Trusted by Defenders”. The name is a joke about the model’s size, and the size is the story: Large 4 is a natively multimodal mixture-of-experts model with about one trillion total parameters, of which 52 billion are active for any given token, plus a 1.6-billion-parameter vision encoder. It is, by Mistral’s account, the largest and most capable model the company has ever built, and the first flagship “Large” upgrade since Large 3 in December 2025.

The training story is part of the pitch. Large 4 was trained from scratch on 3,800 Nvidia Grace Blackwell GPUs in Mistral’s own datacentres in Europe, and the public preview was served from that same infrastructure. A significant share of the training data was multilingual, spanning more than 160 languages, including every official language of the European Union. Mistral’s reinforcement-learning operation runs at a scale it now talks about openly: around 3,000 GPUs in training, generating roughly 33 billion tokens a day, of which about 16 billion are trainable completion tokens after filtering. The point of publishing those numbers is to show that the model was not assembled from someone else’s weights or trained on rented infrastructure in another jurisdiction. For the sovereignty customers Mistral is courting — governments, banks, defence and healthcare organisations — where a model was trained and where it runs are part of the product, not decoration. That is also why Mistral makes a point of the model’s European deployment being operated end to end by Mistral itself, under European law, independently of other digital service providers.

What Large 4 is actually for, in Mistral’s telling, is two jobs above all: agentic coding and cybersecurity. On coding, Mistral reports that three weeks of additional reinforcement learning after the October preview took the model to the lead among open-weight models on FrontierCode 2.0, a benchmark that asks whether a repository maintainer would merge an agent’s change, with a score of 19.7%, and on Terminal-Bench 4.0, where it improved by 13 percentage points to 41% while cutting output tokens by 30%. Its combined score on the Artificial Analysis Coding Agent Index is 48%, effectively tied with the leading open-weight model, GLM-5.3, while using 47% fewer tokens per task. In a blind human evaluation run by Surge AI, professional annotators rating code quality on a one-to-five scale ranked Large 4 second of five models at 3.74, ahead of Kimi K3, GLM-5.3 and GLM-5.2, and behind only Claude Opus 5 at 4.22. On general agentic work it scores 59.9% on AutomationBench, a set of 657 business workflows across Gmail, Google Sheets, Slack and Salesforce, and 1,393 Elo on AA-Briefcase, which tests long-horizon knowledge work. All of these figures are Mistral’s own reports, from its own announcements and evaluations it commissioned; they are not independent verification, and this guide presents them as such.

The cybersecurity case is where Large 4 is most distinctive, and where Mistral’s open-weight strategy has its sharpest edge. On the Artificial Analysis Cyber Index, which measures a model’s ability to find vulnerabilities in a codebase, reproduce them and patch them without breaking existing functionality, Large 4 scores 51%. On Cybench, a set of 40 exercises drawn from security competitions, it solves 93% of the challenges. On one test in the Cyber Index — reproducing a real vulnerability in open-source software and then patching it — it scores 82%, the highest of any model. Mistral’s argument is that this work is precisely what closed models often refuse to do: several leading closed models, including Claude Opus 5.5 and GPT-6 Astra, score near zero on the same reproduction test because they refuse the task, while defenders need to prove a flaw is real before they can fix it. At the same time, Mistral reports that the model’s guardrails distinguish defensive from malicious cyber work, refusing an average of 96.6% of malicious cyber requests across JailbreakBench, StrongREJECT and AgentHarm — the highest refusal rate it measures among open models — and resisting 97.1% of attacks on the B3 Agent Security Benchmark, up from 93.3% in the preview. Whether that balance holds in hostile hands is a question no benchmark can settle, and Mistral itself notes that it has seen potential for the model to uplift real attacks in live traffic and is adjusting its guardrails accordingly. A specialised variant optimised for penetration testing, vulnerability assessment and red-teaming is available in preview to vetted enterprise design partners, including Orange Cyberdefense, under contractual safeguards and restricted to owned or explicitly authorised systems.

Large 4 is also a genuinely multimodal model in a way its predecessors were not. It reasons across complex documents, charts and natural images, and it can combine visual grounding with agentic capabilities — inspecting gigapixel satellite imagery, verifying mechanical parts in technical drawings, retrieving evidence from PDFs, and scanning geospatial images for hard-to-find objects. On visual grounding specifically, Mistral reports that it surpasses GPT-6 Astra on Dense 200, 42% against 41%. Its science and mathematics capabilities are built for agentic coding on scientific tasks: Mistral says it is state of the art among open-weight models on SciCode-Verified, and that it can generate a full Hartree–Fock chemistry simulation in a single attempt. On knowledge work — spreadsheets, documents, legal and financial tasks — it was evaluated by the third-party firm vals.ai and, Mistral reports, exceeded GPT-6 Astra on representative legal and financial tasks, and outperformed all open-source models on HarveyAI’s Legal Agent benchmark. Its web-browsing score on BrowseComp reached 84.5% after reinforcement learning, ahead of GLM-5.3 and DeepSeek’s V4.1 Flash, according to Mistral.

Three practical details matter for anyone considering Large 4. The first is the reasoning modes: the model now supports three settings — Low for simple tasks, Medium for everyday tasks, and High, the default, for the hardest work requiring precision and thorough reasoning — so developers can trade performance against cost instead of paying for maximum reasoning on every call. The second is the price: at general availability it costs $0.70 per million input tokens, $0.07 per million cached input tokens, and $2.10 per million output tokens. That is roughly half the preview pricing of $1.36 and $4.18, which Mistral charged during the October preview. The third is where it runs: through Mistral’s own Studio and API, through OpenRouter, through Microsoft Foundry — Microsoft was the first cloud provider to offer it — and through Nvidia’s hosted endpoints and NIM containers, or self-hosted from the open weights on Hugging Face using vLLM or SGLang. The weights are free to use, modify, fine-tune and commercialise under the Mistral Ecosystem Open-Weight licence, a modified MIT licence, with one condition: any business that operates a Model-as-a-Service offering must obtain a separate commercial licence from Mistral first.

Mistral Small 4: The Hybrid Workhorse

Three interlocking glowing gears inside one compact machine, a hybrid model doing three jobs. No text.

If Large 4 is Mistral’s flagship, Small 4 is the model most people will actually use. Announced on 16 March 2026, it is the first Mistral model to unify the capabilities of three previously separate lines — Magistral for reasoning, Pixtral for multimodal understanding, and Devstral for agentic coding — into a single, versatile model. The idea is simple and, for Mistral, a genuine change of direction: instead of asking users to choose between a fast instruct model, a powerful reasoning engine and a multimodal assistant, one model now delivers all three, with a setting that adjusts how hard it thinks. It is the model that makes the retirements of Magistral and Devstral make sense. They did not disappear; they were folded into Small 4.

The architecture is a mixture of experts: 128 experts, with four active per token, giving 119 billion total parameters and about 6.5 billion active per token, according to Mistral’s documentation. (The launch announcement describes it slightly differently, as 6 billion active per token, or 8 billion including the embedding and output layers; this guide uses the documentation’s figure.) It accepts both text and image inputs, and it has a 256,000-token context window — large enough for long documents and substantial codebases. It is released under the Apache 2.0 licence, the most permissive licence Mistral uses, which means it can be used, modified and deployed commercially with almost no restrictions. That licence choice is deliberate: Small 4 is the model Mistral wants everywhere, including on hardware it does not control, and Apache 2.0 is the licence that lets that happen.

The feature that defines Small 4 is the reasoning_effort parameter. Set it to “none” and the model answers fast, in the chat style of the earlier Small 3.2. Set it to “high” and it reasons step by step, with verbosity equivalent to the old Magistral models. In between, it scales. That single control replaces what used to be a model-selection decision, and it changes the economics of running the model: simple questions cost almost nothing in time and tokens, while hard questions get the full reasoning treatment. Mistral’s performance claims for the model are framed in exactly those economic terms. It reports a 40% reduction in end-to-end completion time in a latency-optimised setup, and three times more requests per second in a throughput-optimised setup, compared with Mistral Small 3. On the Artificial Analysis long-context reasoning benchmark it scores 0.72 with just 1,600 characters of output, while Qwen models need three and a half to four times more output for comparable performance. On LiveCodeBench it outperforms GPT-OSS 120B while producing 20% less output. Shorter outputs mean lower latency, lower inference costs and a better user experience — and Mistral is explicit that this efficiency gap is the point.

Small 4 is also designed to be run by its users, not just rented from Mistral. The minimum infrastructure Mistral specifies is four Nvidia HGX H100 GPUs, two HGX H200s, or a single DGX B200, with a recommended setup of four H100s, four H200s or two DGX B200s for optimal performance. It is available on the open inference stacks — vLLM, llama.cpp, SGLang and Transformers — and was available on day one as an Nvidia NIM, the containerised inference package, with customisation through Nvidia’s NeMo framework. Developers can also prototype it for free on Nvidia’s build platform. Through Mistral’s own API it is the cheap option in the current lineup: $0.15 per million input tokens and $0.60 per million output tokens, against Large 4’s $0.70 and $2.10. That price gap is the practical reason Small 4 matters. For bulk work — chat assistants, document understanding, multimodal analysis, coding automation, codebase exploration — it is the model Mistral itself recommends when cost and latency matter more than maximum capability, and the pricing page’s FAQ names Mistral Small for exactly those cost-sensitive projects.

The intended audience, as Mistral describes it, is broad: developers for coding automation and agentic workflows; enterprises for general chat assistants, document understanding and multimodal analysis; researchers for mathematics, research and complex reasoning. Its open-source licence and customisable architecture make it suitable for fine-tuning and specialisation, and Mistral presents it as the model that brings open-source AI to real-world use cases. There is a fair claim in that, and a limit worth stating. Small 4 is genuinely capable for its size, and its efficiency numbers are impressive, but it is not a frontier model. It is the efficient middle of Mistral’s range, and its job is to be good enough, cheap enough and open enough to run everywhere. Judged on that job, it is the most important model Mistral currently ships — the one that will decide whether the company’s open strategy works in production, rather than only in announcements.

Mistral 3 and Ministral 3: The Previous Generation

A phone, a laptop and a small edge device on a desk, each glowing with a small AI orb, models at every size. No text, no logos.

Before Large 4, there was Mistral 3. Announced on 2 December 2025, it was the previous generation of Mistral’s open models, and it is still the generation many existing deployments run on — which is why it deserves a section of its own, even though its flagship is now deprecated. Mistral 3 came in two parts: a large model, Mistral Large 3, and a family of three small dense models called Ministral 3, in 3-billion, 8-billion and 14-billion-parameter sizes. All of them were released under the Apache 2.0 licence, in both base and instruction-tuned versions, and all of them were multimodal, understanding images as well as text.

Large 3 was a sparse mixture-of-experts model with 41 billion active parameters and 675 billion total parameters, trained from scratch on 3,000 Nvidia H200 GPUs. Mistral described it as its first mixture-of-experts model since the seminal Mixtral series, and as a substantial step forward in its pretraining. After post-training it achieved parity with the best instruction-tuned open-weight models on general prompts, while demonstrating image understanding and, Mistral claimed, best-in-class performance on multilingual conversations outside English and Chinese. At launch it debuted at number two in the open-source non-reasoning category of the LMArena leaderboard, and sixth among open-source models overall. It was also built to be deployable: working with Nvidia, vLLM and Red Hat, Mistral released an NVFP4 checkpoint that lets the model run efficiently on Blackwell NVL72 systems and on a single node of eight A100 or eight H100 GPUs.

The Ministral 3 family was the more interesting half of the release, because it was aimed at the edge — the devices and small servers where a trillion-parameter model will never run. Three sizes, 3B, 8B and 14B, each available in base, instruct and reasoning variants, each with image understanding, all under Apache 2.0. The reasoning variants were the surprise: the 14-billion-parameter reasoning model scored 85% on AIME 2025, the American Invitational Mathematics Examination benchmark that has become a standard test of mathematical reasoning. Mistral’s claim for the Ministral line was about cost-to-performance ratio rather than raw capability: the instruct models matched or exceeded comparable models while often producing an order of magnitude fewer tokens, which matters in the real world, where both the number of generated tokens and the model size determine what a deployment costs. The family was designed to run from datacentres down to edge hardware — Nvidia’s DGX Spark, RTX PCs and laptops, and Jetson devices — with optimised inference support through TensorRT-LLM and SGLang.

The status of this generation on 21 November 2026 is mixed, and it illustrates Mistral’s churn problem. Large 3 is on Mistral’s deprecated list: deprecated on 31 October 2026, with Large 4 named as its replacement, and scheduled to leave the API at the end of April 2027. It lived as the flagship for eleven months. The Ministral 3 family was not on the deprecated list when this guide was researched, but its live API status was not re-verified on Mistral’s documentation pages either, so this guide describes it only on the strength of its December 2025 launch announcement. What is certain is the pattern: Mistral’s generations turn over fast, and a model that was the flagship in December 2025 is a legacy system by November 2026. For users, the lesson is to treat the Ministral 3 family as a capable, openly licensed set of small models that may or may not be served by Mistral’s API in the future — and to check the documentation before building on them. For Mistral, the lesson is one the company has not yet fully learned: rapid progress is a selling point for a research lab, but it is a cost for the customers who have to migrate.

Magistral: The Reasoning Line That Became a Feature

A winding path of glowing stepping stones through a maze, step-by-step reasoning made visible. No text.

Magistral was Mistral’s first reasoning model, announced on 10 June 2025, and it is the clearest example of how fast Mistral’s catalogue moves — and how a whole product line can become a feature of another model. The idea behind Magistral was straightforward. General-purpose models answer quickly and confidently; reasoning models think first. They work through a problem step by step, showing their chain of thought, which makes them better at multi-step logic and easier to audit, because every conclusion can be traced back through its logical steps. Mistral’s pitch was that existing reasoning models lacked specialised depth for domain-specific problems, offered limited transparency, and reasoned inconsistently in the user’s language. Magistral was designed to fix all three: domain-specific reasoning, transparent reasoning you can follow and verify, and reasoning that works natively across languages.

It came in two versions. Magistral Small was a 24-billion-parameter open-weight model released under the Apache 2.0 licence, free for the community to download, examine, modify and build on. Magistral Medium was the more powerful enterprise version, available through the API and on cloud marketplaces. On AIME 2024, the mathematics benchmark, Magistral Medium scored 73.6%, rising to 90% with majority voting across 64 attempts; Magistral Small scored 70.7%, rising to 83.3%. Those were strong numbers for mid-2025, and Mistral supported the release with a research paper covering its evaluations, training infrastructure, reinforcement-learning algorithm and observations on training reasoning models. The model was designed to reason in the user’s language across English, French, Spanish, German, Italian, Arabic, Russian and Simplified Chinese, and in Le Chat it shipped with a Think mode and a Flash Answers feature that Mistral claimed delivered responses up to ten times faster than most competitors.

The use cases Mistral imagined for Magistral were the ones where traceable reasoning matters most. In business strategy and operations, it was meant for research, strategic planning, operational optimisation and data-driven decision making. In regulated industries — legal, finance, healthcare and government — its traceable reasoning was supposed to meet compliance requirements, because every conclusion could be traced back through its logical steps. In systems, software and data engineering, it was meant to improve project planning, backend architecture, frontend design and data engineering through sequenced, multi-step actions involving external tools. And in content and communication, Mistral’s early tests found it an excellent creative companion for writing and storytelling. It was, in short, a model for work where being right matters more than being fast, and where showing your work matters as much as the answer.

And then, within sixteen months, it was gone as a standalone product. Magistral Medium 1.2 and Magistral Small 1.2 are on Mistral’s deprecated list, and the Medium 1.2 version left the API on 31 July 2026. The capability did not disappear — it was absorbed. Small 4’s reasoning_effort parameter delivers what Magistral delivered, with verbosity equivalent to the old Magistral models at its highest setting, and Large 4’s three reasoning modes do the same at the flagship level. That is a sensible engineering decision: reasoning is a mode, not a model, and maintaining a separate reasoning line alongside a general line doubles the work for little benefit. But it is also a warning. A team that built a product on Magistral Medium in the summer of 2025 had about a year before it needed to migrate, and the migration was not to a new Magistral but to a different model with a different interface. The name Magistral still matters for understanding Mistral’s history — it was the company’s first serious reasoning model, and the research behind it feeds everything that followed — but as a product you can use today, it is history. This guide keeps the name in the title’s history, not in its present tense, for exactly that reason.

Codestral and Devstral: The Coding Line

A developer's desk at night, a monitor with abstract code-like lines and a soft holographic helper beside it, AI pair programming. No readable words.

Mistral’s coding models have the longest and most tangled history in the catalogue, because coding was the first place the company tried to sell a complete stack rather than a single model. The story starts with Codestral, a model built specifically for code generation and, above all, for fill-in-the-middle completion — the job of suggesting the code that belongs between the lines a developer has already written. The first version, Codestral 25.01, was released in January 2025 and retired in November 2025. The current version, Codestral 25.08, was announced on 30 July 2025 and is still generally available: a 128,000-token-context model priced at $0.30 per million input tokens and $0.90 per million output tokens. Mistral reports that 25.08 delivers 30% more accepted completions, 10% more retained code after suggestion, and 50% fewer runaway generations than its predecessor, with improvements validated in live IDE usage across production codebases, plus better instruction following and code abilities in chat mode. Those are Mistral’s own figures, and they describe a model optimised for the unglamorous reality of professional coding: latency-sensitive, context-aware, self-deployable, and good at the small completions that make up most of a developer’s day.

Devstral was the ambitious half of the coding line: not a completion model but an agentic coding model, built to explore codebases, orchestrate changes across multiple files, track framework dependencies, detect failures and retry with corrections — the work of a junior engineer rather than an autocomplete. The first Devstral, released in May 2025, was a 24-billion-parameter open model. Devstral Small 1.1 scored 53.6% on SWE-Bench Verified, the benchmark that tests whether a model can resolve real GitHub issues, and Devstral Medium reached 61.6%, which Mistral reported as outperforming Claude 3.5, GPT-4.1-mini and other open models by wide margins. Then came Devstral 2, announced on 9 December 2025 with the Vibe CLI: a 123-billion-parameter dense model with a 256,000-token context window, scoring 72.2% on SWE-Bench Verified, released under a modified MIT licence, with a 24-billion-parameter Devstral Small 2 at 68.0% under Apache 2.0. Mistral claimed Devstral 2 was up to seven times more cost-efficient than Claude Sonnet at real-world tasks, and human evaluations run through the Cline agent scaffold showed it clearly ahead of DeepSeek V3.2, though still significantly behind Claude Sonnet 4.5 — a gap Mistral itself acknowledged.

The coding stack around these models was, for a time, a complete system. Codestral Embed provided code-specific semantic search, claimed by Mistral to outperform leading embedding models from OpenAI and Cohere on real-world code retrieval, with configurable output dimensions and fully private deployment. Devstral provided the agentic workflows. And Mistral Code, a plugin for JetBrains and VS Code IDEs, surfaced all of it in the editor: inline completions from Codestral, one-click automations like “write commit message” and “fix function” powered by Devstral, context from Git diffs and terminal history, and semantic search backed by Codestral Embed — with no mandatory telemetry, single sign-on, audit logging and usage controls for enterprise deployment. It was adopted, Mistral reported, by Capgemini across global delivery teams, by the Spanish bank Abanca in a fully self-hosted deployment, and by SNCF, the French national railway, to modernise legacy Java systems with human oversight. The pitch was that enterprises did not want an isolated assistant; they wanted an integrated system that matched the complexity, security posture and velocity of modern software development.

That stack has now been largely superseded, and the coding line’s history is a case study in Mistral’s pace. Mistral has also bought its way into code trustworthiness, acquiring SkipLabs in November 2026 in a move aimed at making AI-written code verifiable. Devstral 2 was retired from the API on 31 July 2026, after about eight months. Mistral Medium 3.5 was itself deprecated on 31 October 2026, with Large 4 as its replacement. And the interface has changed too: the coding agent is now Vibe, in the terminal, in VS Code, on the web and on the desktop, and secondary reporting in November 2026 says the JetBrains plugin Mistral Code is deprecated, though Mistral’s own pages read for this guide do not confirm that. What survives intact is Codestral 25.08, still the completion model, and the idea behind the stack: that coding assistance is most valuable when completion, search and agents share architecture, context and governance, and when the whole thing can run inside the customer’s own infrastructure. The names have changed three times in eighteen months. The idea has not.

Open vs Hosted: What “Open-Weight” Actually Means

Two doors side by side: one open door leading to an open field, one closed door with a keyhole, the open-vs-hosted choice. No text.

The most important word in Mistral’s vocabulary is “open”, and it is also the most misunderstood. Mistral uses it in two different senses, and the difference decides what you can actually do with a model. The first sense is open source, in the traditional meaning: the model is released under the Apache 2.0 licence, which lets anyone use, modify and distribute it, including commercially, with almost no restrictions. Mistral 7B, Mistral Small 4, Magistral Small, Devstral Small 2 and the whole Mistral 3 family were released this way. The second sense is open-weight: the model weights are published and downloadable, so you can inspect them, fine-tune them and run them yourself — but the licence attached to them carries conditions. Large 4 uses Mistral’s own Ecosystem Open-Weight licence, a modified MIT licence, which is free to use, modify, fine-tune and commercialise, with one condition: if you operate a Model-as-a-Service offering — that is, if you sell access to the model itself as a hosted service — you must obtain a separate commercial licence from Mistral first. Medium 3.5 and Devstral 2 used a modified MIT licence on similar terms. The distinction matters. Apache 2.0 means the model is yours. The open-weight licence means the model is yours to run, but not yours to resell as a service.

Why does Mistral publish weights at all, when its competitors keep theirs secret? The company’s founding argument is that openness is a better business, not just a better philosophy. With weights in hand, developers can fine-tune a model on their own data, customise its guardrails and tone, deploy it on their own infrastructure, and audit it for flaws. Enterprises can keep sensitive data inside their own walls, which is the core of the sovereignty pitch. And researchers can inspect, adapt and build on the work, which is how open ecosystems improve. Mistral also argues, in its Large 4 announcement, that openness is a safety feature for defenders: a security team that can run a model on its own infrastructure, under its own policies, is not exposed to a provider changing its refusal behaviour in the middle of an incident, or withdrawing a capability it depends on. That argument is self-interested — Mistral is selling the alternative — but it is not wrong. The history of hosted AI is full of models that changed behaviour overnight, and a team that owns its weights is insulated from that.

The hosted side of the business is the API, now sold through Mistral Studio, the console at console.mistral.ai that replaced the old La Plateforme. Through the API you can call any current Mistral model, plus — since 2026 — other vendors’ models that Mistral resells, including Z.ai’s GLM 5.3 and DeepSeek’s V4.1 Flash. The API is OpenAI-compatible, offers batch processing at a 50% discount and cached input tokens at up to 90% off, and is available through global, European and US endpoints, with regional endpoints for customers who need processing to stay in a particular jurisdiction. You can also get Mistral models through the cloud marketplaces: Large 4 is available through Microsoft Foundry, Amazon Bedrock and SageMaker, Azure AI, Google Cloud Marketplace, IBM WatsonX, OpenRouter, Fireworks, Together AI and Modal, among others. And you can self-host, from the open weights, using the open inference stacks vLLM and SGLang, or as a containerised Nvidia NIM. The choice is not between open and hosted. It is between three ways of running the same model — Mistral’s API, someone else’s cloud, or your own machines — and the right answer depends on what you care about most: convenience, cost, or control.

There are limits to the openness, and an honest guide should name them. First, open-weight is not open source in the full sense: Mistral does not publish its training data, and the training recipes are described in announcements and technical reports rather than released as reproducible code. You can inspect the model, but you cannot fully reproduce how it was made. Second, the newest and most capable models are the ones with the most restrictive licences — Large 4’s open-weight licence, not Apache 2.0 — while the fully permissive licences attach to the smaller models. That is a commercial decision, and an understandable one, but it means the openness is graded by capability. Third, openness does not solve the hardware problem. A trillion-parameter model needs serious infrastructure to run: even the efficient Small 4 needs four high-end GPUs at minimum, and Large 4 needs far more. For most users, “open” will in practice mean “available through an API I can switch away from”, rather than “running on a server I own”. That is still a meaningful difference from a closed model — you can leave, you can fine-tune, you can audit — but it is not the same as every user running their own frontier model. Mistral’s own language reflects this: it talks about sovereignty as control across four dimensions — data that stays inside the organisation, models that are controllable and customisable, compute that is private and predictable, and production systems that are fully controllable and auditable. Open weights are the foundation of that control, not the whole building.

Vibe: One Agent for Work and Code

One agent juggling floating panels — inbox, calendar, documents and code windows — in a single workspace. No readable words.

If the models are Mistral’s engine, Vibe is its steering wheel — the product most people will actually touch. Vibe is Mistral’s AI agent for work and coding, and until 28 May 2026 it was called Le Chat. The rename was more than a rebrand. Le Chat was a chatbot, a place to ask questions. Vibe is an agent: it does multi-step work on your behalf, across your inbox, calendar, documents and code, and it keeps working while you do something else. Mistral’s own summary is that Vibe is “one agent for long-running, multi-step work”, and the product is organised around exactly two modes: Work, for office tasks, and Code, for software development. One licence covers both, and everyone who was on Le Chat had their conversations, settings and plans carried over automatically.

Work Mode is the office half. It catches up across your inbox and calendar, runs deep research, drafts deliverables, and orchestrates the recurring processes of a working day. It grounds its work in your context through connectors — Google Workspace, Outlook, SharePoint, Slack, GitHub, and custom connectors or document libraries — and it can search enterprise knowledge, analyse structured data from a database or a spreadsheet, render charts and dashboards inside the conversation, and draft documents, reports, RFP responses and board decks using a Canvas tool, ready to push to Notion, SharePoint or your inbox. Tasks can be scheduled to run once or on a daily, weekly or monthly cadence, with a notification when a run finishes. Reusable Skills, built on open standards, automate repeatable workflows. And every step is visible: each tool call and reasoning chain can be expanded to show its inputs and outputs, so you can see what the agent did and why. For one person, Mistral says, this turns a morning of admin into a single prompt. For an organisation, the same agent runs the processes that keep a business moving, governed by the permissions set at admin level.

Code Mode is the software half, and it is where Mistral has invested most heavily in 2026. It takes coding work from request to merged change: you connect to GitHub, start a session from the web app, and the agent builds features, fixes bugs, refactors code and ships reviewable pull requests, running in an isolated sandbox while you manage sensitive actions and inspect diffs. Sessions can run in parallel, persist while your machine is off, and be triggered from third-party apps. The same agent is available in the terminal through the Vibe CLI — the open-source command-line coding assistant introduced with Devstral 2 in December 2025, which explores, modifies and executes changes across a codebase using natural language — and in the editor through a VS Code extension that works across a whole project in a side panel, reading, editing and executing commands beside your files, with open files attaching automatically and @ mentions pulling in context from other directories. On 17 November 2026, Mistral added a desktop app for macOS and Windows, Vibe Desktop, which brings Work, local coding and cloud coding sessions together in one window, letting you hand several tasks to agents at once, watch them work side by side, and review their diffs and pull requests in the same place. A Linux version is listed as coming soon. The remote coding agents in Vibe were powered by Mistral Medium 3.5 — now deprecated, with Large 4 as its documented replacement — and the underlying models are, Mistral says, flagship models optimised for reasoning, agentic tasks, tool calls and coding.

Vibe is priced as a subscription, in four tiers. The Free plan covers quick answers and simple everyday tasks, with limited messages, web searches and coding sessions, basic image generation and editing, and access to more than 100 connectors. Pro costs $14.99 a month and offers up to six times the free plan’s messages and web searches, faster responses, all-day coding in the CLI, IDE or on the web, spreadsheets and slides generation, better image generation and editing, and chat and email support. Students can get Pro for $5.99 a month. Team costs $24.99 per user per month and adds a shared workspace with admin controls, up to 30 gigabytes of storage per user, domain name verification and data export. Enterprise is custom, with private deployments, custom models, custom agents and workflows, audit logs, single sign-on and white labelling. One detail worth noting: Mistral says it will soon sunset Chat mode, the direct descendant of Le Chat, with all history preserved in Work Mode. The chatbot is becoming an agent, and the company is not keeping the old interface around. That is a confident move, and a slightly ruthless one, but it matches the direction of the whole product: fewer places to chat, more places to get work done.

Mistral Studio: The Enterprise Platform

A mission-control wall of dashboards, pipelines and charts, an enterprise AI platform. No readable words.

Models and an assistant are not enough for a large organisation. What enterprises actually need, Mistral argues, is a system that turns AI from a collection of prototypes into a reliable, observable and governed capability — and that is what Mistral Studio is for. Announced on 24 October 2025, Studio is the company’s production AI platform, and it is the product that ties the rest of the catalogue together. Its pitch starts from an observation that will be familiar to anyone who has worked in enterprise IT: teams have built dozens of prototypes — copilots, chat interfaces, summarisation tools, internal question-answering — but most AI adoption stalls at the prototype stage, because teams cannot track how outputs change across model or prompt versions, reproduce results, monitor real usage, run evaluations tied to their own benchmarks, fine-tune on proprietary data, or deploy governed workflows that satisfy security and compliance constraints. Studio is Mistral’s answer to that gap, built from the infrastructure the company uses to operate its own systems at scale.

The platform is organised around three pillars. The first is Observability: full visibility into what is happening, why, and how to improve it. An Explorer lets teams filter and inspect traffic, build datasets and identify regressions. Judges — evaluation logic that can be built and tested in a Judge Playground — score outputs at scale. Campaigns and Datasets convert production interactions into curated evaluation sets automatically. Experiments, iterations and dashboards make improvement measurable rather than anecdotal. The idea is to trace outcomes back to prompts, prompts back to versions, and versions back to real usage, closing the feedback loop with data instead of intuition. The second pillar is the Agent Runtime: the execution backbone that runs every agent, from simple single-step tasks to complex multi-step business flows, with durability, transparency and reproducibility. Each agent operates inside a stateful, fault-tolerant runtime built on Temporal, which guarantees consistent behaviour across retries, long-running tasks and chained calls, manages large payloads, offloads documents to object storage, and generates static graphs that make execution paths auditable. Every execution emits telemetry and evaluation data that flows back into Observability. The third pillar is the AI Registry: a system of record for every asset across the AI lifecycle — agents, models, datasets, judges, tools and workflows — tracking lineage, ownership and versioning end to end, enforcing access controls, moderation policies and promotion gates before deployment, and making every asset discoverable, auditable and portable across environments.

On top of those pillars sit the products that enterprises actually buy. Managed agents and apps are in public preview, and Workflows — the durable, multi-step business processes that run on the Agent Runtime — have reached general availability. Forge, announced in March 2026, is the system for building frontier-grade models grounded in a customer’s proprietary knowledge: custom training, fine-tuning and fully adapted models, built with Mistral’s Applied AI experts working alongside customer teams. The Large 4 announcement describes the same offer in plainer terms: customers tailor the model with proprietary security data, policies, detection logic and threat environments using Forge, and for well-defined tasks it can be further optimised into smaller, task-specific variants. The platform supports hybrid, dedicated and self-hosted deployments, so enterprises can run agents wherever their infrastructure requires while keeping the same durability, traceability and control. Usage is tracked in the Studio dashboard, including metrics for calls, endpoints and user activity, with costs broken down by service — completions, OCR, fine-tuning — so teams can monitor and optimise consumption.

Studio is also where Mistral’s sovereignty pitch becomes concrete rather than rhetorical. A European bank that needs its data to stay in Europe, its models to be auditable, and its deployments to survive a vendor changing its mind can, in principle, get all three from Studio: European-hosted inference, open weights it can inspect and run itself, and a registry that records exactly which version of which prompt and model produced which output. The site’s own reporting has covered the enterprise side of this story — the managed agents and apps preview, and the partnerships with Deutsche Telekom and Orange that put Mistral’s models inside European infrastructure. But it is worth being clear about what Studio is not. It is not a product for individuals, and it is not cheap or simple to adopt. It is a platform for organisations that already have AI in production and are struggling to operate it reliably. For those organisations, the question is not whether Mistral’s models are the best in the world — on most benchmarks they are not — but whether the combination of capable-enough models, open weights, European hosting and production tooling is worth more than a marginally better closed model that they cannot control. That is the bet Studio embodies, and it is the bet the company’s future depends on.

Building with the API: Agents, Search, OCR and Voice

Building blocks and tools assembling an agent from pieces — search, code, documents, voice. No text.

For developers, Mistral’s offering is the API, and the API has been changing as fast as the models. The headline change, visible in Mistral’s documentation in November 2026, is that the Agents API version 1 is being deprecated, and new agents should be built with the Agents SDK — a Python and TypeScript library in which an agent is an object in your own code, its harness runs the loop, calls the model, runs your tools, and records each step in a session, and you deploy it with Studio, self-hosted or Mistral-managed. The old pattern, in which you created an agent on Mistral’s servers, started a conversation, ran each function the model called and sent the result back, still works for existing users, but Mistral’s documentation is explicit that new work should use the SDK. That is the third interface change in the coding and agents area in eighteen months, after the move from Mistral Code to Vibe and from the Conversations API to the Agents API, and developers should expect the pattern to keep evolving.

What the agent can do, in either interface, is defined by its tools. Mistral provides four built-in tools that run on its own servers and are ready out of the box: Web Search, which searches the web for up-to-date information not in the model’s knowledge, with a premium tier; Code Interpreter, which runs code and generates plots in a sandbox, useful for data analysis and visualisation; Image Generation, which generates images from a prompt; and Document Library, which searches through documents uploaded to your libraries, enabling retrieval-augmented generation from your own information. Alongside those are function calling for your own local tools, and Connectors, which register Model Context Protocol servers as managed connectors so the model can discover and call your internal systems automatically. Agents can be handed off to one another, so a general agent can transfer a conversation to a specialist, and conversations are stateful, persisting across turns. There is also a Responses API, a simpler interface for building agents that this site covered when it launched, for developers who want the built-in tools without the full agent machinery.

Two specialist capabilities deserve attention because they show where Mistral is heading beyond chat. The first is search. Mistral’s Search Toolkit, in public preview, is an open-source, composable framework for production search pipelines — the retrieval layer that helps AI systems navigate, read and verify information inside and across complex documents. Its ingestion component is OCR 4, which means the document understanding and the search system are designed to work together: OCR 4 turns a document into structured, citation-ready blocks, and the Search Toolkit retrieves and evaluates them. The second is voice. Voxtral 3, launched on 22 October 2026, is a family of speech models plus a voice agent SDK built on LiveKit that runs listening, reasoning, tool calling and speaking as a single real-time streaming pipeline. The speech-to-text model covers more than 50 languages with diarization and context biasing; the text-to-speech model covers more than 35 languages and 50 dialects, with emotion, pacing and emphasis control and voice cloning from 30 minutes of audio; and the realtime transcription model works at sub-500-millisecond latency in more than 35 languages. During a call, an agent can verify a caller’s identity, check a bill or book an appointment, using function calling, web search and retrieval over your own systems, over SIP, WebRTC or existing telephony and contact-centre software.

The document side is Mistral OCR 4, launched on 23 June 2026 — OCR 4.1 is the current version as of 21 November 2026, per Mistral’s model list — and it is the most enterprise-ready of the specialist models. It extracts and structures content from PDFs, Word documents, PowerPoint and OpenDocument files in 170 languages, returning not just text but bounding boxes that localise each element, block classifications that say whether a block is a title, a table, an equation or a signature, and confidence scores per page and per word. That structure supports semantic chunking for retrieval, agents that fill in forms and process invoices, and structured data pipelines in which confidence scores tell human verifiers where to look. It runs in a single container for fully self-hosted deployment, and it is available through the API at $4 per 1,000 pages, or $2 with the Batch API, with a no-code Document AI option at $5 per 1,000 pages that returns structured JSON against a schema you define. Mistral reports that independent annotators preferred OCR 4 over every leading OCR and document-AI system tested, with win rates averaging 72%, and that it scored 85.20 on OlmOCRBench and 93.07 on OmniDocBench — but the company itself warns that both benchmarks have known scoring limitations, that most of the mismatches it audited were artefacts of how the benchmarks compare output rather than model errors, and that the aggregate scores should be treated as directional. That candour is refreshing, and it is the right attitude to bring to every benchmark in this guide.

What Mistral Costs

A neat market stall with piles of coins and tokens, pricing made simple. No numbers, no readable text.

Mistral’s pricing is simple to describe and easy to get wrong, because the numbers have changed as fast as the models. The prices in this section are the ones in Mistral’s own model documentation and announcements on 21 November 2026, and where the company’s own pages disagree, this guide says so. The API is priced per million tokens, with input and output counted separately. Mistral Large 4 costs $0.70 per million input tokens, $0.07 per million cached input tokens, and $2.10 per million output tokens. Mistral Small 4 costs $0.15 per million input tokens and $0.60 per million output tokens. Codestral 25.08 costs $0.30 per million input tokens and $0.90 per million output tokens. The third-party models Mistral resells are priced higher: Z.ai’s GLM 5.3 costs $1.40 per million input tokens and $4.40 per million output tokens, according to Mistral’s model selection guide. One caution: a line in the frequently asked questions on Mistral’s pricing page still quotes $0.50 in and $1.50 out for “Mistral Large”, which conflicts with the $0.70 and $2.10 in the model documentation and the Large 4 announcement. This guide uses the documentation figures, which are the current ones; the FAQ figure appears to be stale.

The specialist models are priced differently, because they are not priced by the token. Mistral OCR 4 costs $4 per 1,000 pages through the API, falling to $2 per 1,000 pages with the Batch API, while the no-code Document AI option costs $5 per 1,000 pages. Voxtral’s speech models are priced per minute or per character: Mini Transcribe 3 at $0.0035 per minute, Small Transcribe 3 at $0.0040 per minute, Realtime Transcribe 3 at $0.008 per minute, Mini TTS 3 at $13 per million characters, and the full voice agent API at $0.04 per minute with Small 4 doing the reasoning or $0.08 per minute with a larger model. Across the API, batch processing reduces prices by 50% for high-volume work, and cached input tokens — repeated prompts — reduce input costs by up to 90%. Those two discounts are where the real savings are: a team that batches its document processing and caches its repeated system prompts will pay a fraction of the headline rate, and Mistral’s own pricing page makes the point explicitly.

The assistant, Vibe, is priced as a subscription rather than by the token. The Free plan costs nothing and covers quick answers, simple everyday tasks, limited messages and web searches, limited coding sessions, basic image generation and editing, and more than 100 connectors. Pro costs $14.99 a month and multiplies the free limits — up to six times more messages and web searches, up to forty times more image generations — while adding all-day coding, faster responses, spreadsheet and slide generation, and chat and email support. Verified students can get Pro for $5.99 a month. Team costs $24.99 per user per month and adds a shared workspace with admin controls, up to 30 gigabytes of storage per user, domain name verification and data export. Enterprise is priced on application, with custom deployments, custom models and agents, audit logs, single sign-on and white labelling. Usage beyond the plan limits can be extended with pay-as-you-go credits.

How does that compare with the alternatives? On price alone, Mistral is competitive rather than cheapest. Large 4 at $0.70 in and $2.10 out is substantially cheaper than the flagship closed models, which typically charge several dollars per million input tokens and tens of dollars per million output tokens, but it is not the cheapest open model on the market — DeepSeek’s models and some of the Chinese open-weight models undercut it, and the third-party models Mistral resells through its own API are more expensive than its own. The honest comparison is not the headline rate but the total cost of a workload: the price of the model, the number of tokens it needs to finish a task, and the cost of running it yourself if you self-host. Large 4’s selling point on that measure is token efficiency — it matched GLM-5.3 on the Artificial Analysis Coding Agent Index while using 47% fewer tokens per task, and cut output tokens by 30% between preview and general availability — which means the bill for a finished job can be lower than the per-token price suggests. Small 4’s selling point is the opposite: it is cheap enough per token that bulk work becomes affordable, and open enough under Apache 2.0 that you can run it on your own hardware and stop paying Mistral entirely. The pricing question, like the licensing question, is really a control question: how much are you willing to pay for the right to walk away?

Mistral vs Llama vs Qwen: The Open-Model Comparison

Three distinct coloured paths racing side by side toward the same horizon, an open-model comparison. No text.

Mistral does not compete in a vacuum. The open-model world has three poles: Meta’s Llama, Alibaba’s Qwen, and Mistral itself, with DeepSeek as a fourth force that is cheaper than all of them and more controversial than any of them. The differences between them are not mostly about benchmark scores, which change with every release and which each lab reports in its own favour. They are about strategy, licensing, and who each lab is building for — and those differences are stable enough to write down. This section compares on those grounds only. Current parameter counts, context windows, prices and benchmark scores for Llama and Qwen were not re-verified live for this guide, and this guide does not quote them from memory; the links below go to this site’s own explainers on each family, which carry their own checked dates.

Start with licensing, because it is where the three labs differ most. Qwen is the most permissive of the three: Alibaba releases its models under the Apache 2.0 licence, the same licence Mistral uses for its smaller models, which means almost no restrictions on commercial use. Mistral is in the middle: its smaller models are Apache 2.0, but its flagship Large 4 uses its own Ecosystem Open-Weight licence, which is free to use and commercialise but requires a separate commercial licence if you sell the model as a hosted service. Llama is the most restrictive: Meta’s Llama models are available under a community licence with conditions, including a requirement that very large deployments seek a separate licence from Meta, and the weights are not open source in the sense that Apache 2.0 is. For a developer choosing a model to build a business on, that ordering matters. Qwen gives you the most freedom, Mistral gives you freedom with a condition at the top end, and Llama gives you a capable model with strings attached. None of that makes any of them a bad choice — it makes them different choices, and the right one depends on whether you plan to host the model yourself, sell access to it, or simply call an API.

The second difference is what each lab is for. Meta builds Llama as part of a much larger business: the models power Meta AI across WhatsApp, Instagram and Facebook, and the open release is a way to set standards and attract developers to Meta’s ecosystem. Alibaba builds Qwen as part of a cloud and e-commerce empire: the models are deeply integrated with Alibaba Cloud, and the open release drives usage of the company’s infrastructure. Mistral builds models as its whole business: there is no social network and no cloud empire behind it, only the models, the API, the assistant and the enterprise platform. That focus is both a strength and a weakness. It means Mistral’s incentives are aligned with its users in a way the others’ are not — it has to win on the models and the platform, because there is nothing else. But it also means Mistral is the smallest of the three by an enormous margin, with a fraction of the compute, the headcount and the distribution. Its answer to that disadvantage is the sovereignty pitch: it is the only one of the three that is European, that trains in Europe, that hosts in Europe, and that can credibly sell to a European government or bank on the basis of independence from both American and Chinese technology. Llama cannot offer that, and Qwen cannot offer it either, and for the customers who care about it, that is worth more than a few benchmark points.

The third difference is the shape of the catalogue. Qwen’s family is the broadest and the most confusing: a general Qwen line, a QwQ reasoning line, a Qwen-Coder line, and a long tail of specialised models, with new versions arriving constantly. Llama’s family is the most stable: a small number of sizes, released on a predictable cadence, with the ecosystem — the tooling, the fine-tuning frameworks, the community — that comes from being the default open model for most developers. Mistral’s family is the most volatile: models arrive, merge and retire within months, and the catalogue in November 2026 looks very different from the catalogue of a year earlier. That volatility is the price of Mistral’s speed. It also means the comparison a reader makes today will be out of date by the time the next flagship arrives — which, on Mistral’s current pace, could be six months away. The practical advice is the same as for the catalogue itself: choose on the basis of the model that exists today, pin the exact version, and assume you will migrate. If you want stability and the biggest ecosystem, Llama is the safer bet. If you want permissive licensing and the broadest family, Qwen is the safer bet. If you want European hosting, open weights with a sovereignty story, and a lab whose whole business depends on getting this right, Mistral is the bet — and this site’s own Qwen and DeepSeek explainers are the place to start if you want the other side of the comparison in the same depth.

Who Should Use Mistral — and the Verdict

A lighthouse on a European coast beaming light across the sea, sovereignty and guidance. No text.

After three years, four flagship generations, a dozen retired models and a €3 billion funding round, the question is no longer whether Mistral is serious. It is who Mistral is for. The answer, on the evidence of its current catalogue, is narrower and more interesting than “everyone who wants AI”. Mistral is for the user who needs a capable model and also needs to control it: the European bank that cannot send customer data to an American API, the defence contractor that cannot depend on a provider changing its refusal behaviour mid-incident, the healthcare company that needs to audit exactly what its model does, the developer who wants to fine-tune a model on proprietary code and run it on their own machines, and the enterprise that wants an agent platform it can operate under its own governance. For those users, Mistral’s combination — open weights, European hosting, a full stack from models to Studio, and a price that undercuts the closed flagships — is genuinely distinctive. No other lab offers all four at once.

It is also for the developer who wants a good coding model without paying closed-model prices. Large 4’s coding and cybersecurity numbers, as Mistral reports them, put it at the front of the open-weight field, and Small 4’s efficiency makes bulk coding work affordable. The Vibe agent, with its terminal, editor, web and desktop surfaces, is a serious attempt to meet developers where they work, and the fact that the CLI is open source under Apache 2.0 matters: it is a tool the community can inspect and improve, not just a product to rent. For a small team or an individual developer, the free tier of Vibe and the Apache-licensed Small 4 are a genuinely low-cost way into frontier-capable AI, and the ability to self-host means the cost can fall to hardware alone.

The honest weaknesses belong in the verdict too. The first is churn. Mistral retires models fast — Devstral 2 lasted eight months, Medium 3.5 was deprecated after six, Large 3 after eleven — and every retirement is a migration for someone. A lab that moves this fast is exciting to follow and expensive to build on, and Mistral has not yet shown that it can keep a flagship stable for the two or three years an enterprise deployment needs. The second is capability. On most general benchmarks, Large 4 is competitive with the best open models and behind the best closed ones: its own human evaluation puts it behind Claude Opus 5 on code quality, and its own announcements concede gaps against closed models on some tasks. If you need the single most capable model in the world and you do not care where it runs, Mistral is not your lab. The third is ecosystem. Llama has a larger community, more tooling and more fine-tunes; Qwen has a broader family and the weight of Alibaba Cloud behind it. Mistral’s ecosystem is growing — the Hugging Face downloads, the Nvidia partnership, the cloud marketplaces — but it is smaller, and a smaller ecosystem means more work for the user.

The verdict, then, is a conditional one. Mistral AI is the most interesting open-model lab in the world, and the most important European one, because it has understood something its rivals have not: that for a growing number of customers, control is the product. Open weights, European training and hosting, self-deployment, and a platform that records exactly what your AI did are not features of the model — they are the reason to choose the model. The €3 billion Series D, the trillion-parameter flagship, the agent that does your office work and your coding, and the enterprise platform that governs it all are bets on that idea, and they are bets that have not yet been proven. The models will keep changing — the next flagship is probably six months away, and some of the names in this guide will be retired by the time you read it again. What is unlikely to change is the strategy, because it is the only strategy Mistral has: build the models openly, run them in Europe, and let the customer own the intelligence. For the users who need that, there is no substitute. For everyone else, it is an option worth understanding — and this guide, checked on 21 November 2026, is the map of where that option stands today.

Sources

  • Mistral AI — “Introducing Mistral Large 4” (mistral.ai/news/mistral-large-4/), published 6 October 2026, updated 29 October 2026. Read in full, 21 November 2026.
  • Mistral AI — “Le Chonk: Built for Coders, Trusted by Defenders” (mistral.ai/news/mistral-large-4-oss/), published 29 October 2026, updated 5 November 2026. Read in full, 21 November 2026.
  • Mistral AI — “Making sovereign, open-weight AI the technology frontier” (mistral.ai/news/mistral-makes-sovereign-open-weight-ai-to-frontier/), Series D announcement. Read in full, 21 November 2026; announcement date 8 September 2026 confirmed by secondary reporting (Slator, Le Monde, Business Times, HPCwire).
  • Mistral AI — “Vibe gets to work” (mistral.ai/news/vibe-agent/), published 28 May 2026, updated 5 October 2026. Read in full, 21 November 2026.
  • Mistral AI — “Introducing Mistral Small 4” (mistral.ai/news/mistral-small-4/), published 16 March 2026, updated 8 October 2026. Read in full, 21 November 2026.
  • Mistral AI — “Magistral” (mistral.ai/news/magistral/), published 10 June 2025, updated 8 October 2026. Read in full, 21 November 2026.
  • Mistral AI — “Introducing Mistral 3” (mistral.ai/news/mistral-3/), published 2 December 2025, updated 8 October 2026. Read in full, 21 November 2026.
  • Mistral AI — “Announcing Codestral 25.08 and the Complete Mistral Coding Stack for Enterprise” (mistral.ai/news/codestral-25-08/), published 30 July 2025, updated 8 October 2026. Read in full, 21 November 2026.
  • Mistral AI — “Introducing: Devstral 2 and Mistral Vibe CLI” (mistral.ai/news/devstral-2-vibe-cli/), published 9 December 2025, updated 8 October 2026. Read in full, 21 November 2026.
  • Mistral AI — “Voxtral 3: Speech-to-Text, TTS & Voice Agents” (mistral.ai/news/voxtral-3/), published 22 October 2026. Read in full, 21 November 2026.
  • Mistral AI — “Mistral OCR 4: SOTA OCR for Document Intelligence” (mistral.ai/news/ocr-4/), published 23 June 2026, updated 7 October 2026. Read in full, 21 November 2026.
  • Mistral AI — “Introducing Mistral AI Studio” (mistral.ai/news/ai-studio/), published 24 October 2025, updated 12 November 2026. Read in full, 21 November 2026.
  • Mistral AI — “Bringing open AI models to the frontier” (mistral.ai/news/about-mistral-ai/), published 27 September 2023, updated 8 October 2026. Read in full, 21 November 2026.
  • Mistral AI — About page (mistral.ai/about/), founders, key dates, customers. Read in full, 21 November 2026.
  • Mistral AI — Pricing page (mistral.ai/pricing), Vibe plans and API pricing notes. Read in full, 21 November 2026.
  • Mistral Docs — Models Overview (docs.mistral.ai/getting-started/models/), current and deprecated/retired model lists. Read in full, 21 November 2026.
  • Mistral Docs — Model Selection Guide (docs.mistral.ai/inference/model-selection-guide), prices and specifications. Read in full, 21 November 2026.
  • Mistral Docs — Mistral Large 4 model page (docs.mistral.ai/models/mistral-large-4-0), GA 5 November 2026. Read in full, 21 November 2026.
  • Mistral Docs — Mistral Small 4 model page (docs.mistral.ai/models/mistral-small-4-0-26-03), GA 16 March 2026. Read in full, 21 November 2026.
  • Mistral Docs — Mistral Medium 3.5 model page (docs.mistral.ai/models/mistral-medium-3-5-26-04), deprecated 31 October 2026. Read in full, 21 November 2026.
  • Mistral Docs — Codestral 25.08 model page (docs.mistral.ai/models/codestral-25-08) and Codestral 25.01 model page (docs.mistral.ai/models/codestral-25-01), retired 30 November 2025. Read in full, 21 November 2026.
  • Mistral Docs — Agents Tools Overview (docs.mistral.ai/studio/agents/agent-tools), built-in tools and Agents SDK migration. Read in full, 21 November 2026.
  • Mistral Docs — Release notes (docs.mistral.ai/resources/release-notes), Vibe Desktop general availability 17 November 2026. Read via search result, 21 November 2026.
  • Mistral AI — News index (mistral.ai/news), full catalogue scanned 21 November 2026.
  • Retirement dates (Magistral Medium 1.2 and Devstral 2 leaving the API on 31 July 2026; Medium 3.5 and Large 3 leaving the API at the end of April 2027) — Mistral Docs deprecated-models list confirms the retired status; the dates are corroborated by secondary reporting (vandeblaak.com, 3 November 2026).
  • Site sources linked in the article: the Mistral Large 4 open-weights report (post 3721); the Voxtral 3 report (post 950); the Mistral Studio managed-agents report (post 4550); the Mistral Responses API report (post 3523); the Deutsche Telekom sovereign-AI report (post 6120); the Mistral SkipLabs acquisition report (post 4861); the Qwen, DeepSeek, Llama and NVIDIA explainers.

Share to: , and

Faheem R. Khattak

Senior automation engineer and AI marketing journo covering LLMs, algorithms, analytics, core updates, Bing search, and technical SEO.

0 Comments

Leave a Reply

Optional Take Tag:

🛡️ Spam protection: Honeypot + Time-Trap Active

This site uses Akismet to reduce spam. Learn how your comment data is processed.