Mistral AI

Mistral AI is a French artificial intelligence company that builds large language models and sells access to them. It solves a specific problem for European teams: how to use frontier-grade AI while keeping data inside the EU and, when needed, running the model on your own hardware. Mistral was founded in 2023 in Paris by Arthur Mensch, Guillaume Lample, and Timothée Lacroix. Its distinctive move is a split catalogue - some models ship as open weights under permissive licences, and others stay commercial and API-only.

That split is the whole story, and since December 2025 it has run to three tracks rather than two:

• Apache 2.0 open weights. Mistral Large 3 (announced 2 December 2025), the flagship until October 2026 and still the newest Mistral flagship with published weights, is a sparse mixture-of-experts model of roughly 675B total and 41B active parameters with a 256K context window and image input - and Mistral published both base and instruct weights under Apache 2.0. Mistral Small 4 (16 March 2026) and the Ministral 3 line at 14B, 8B and 3B (2 December 2025) are Apache 2.0 too, as are Voxtral Small, Shieldstral and Leanstral. The older exemplars, Mistral 7B and the Mixtral models, are retired from the hosted API and superseded by Ministral 3.
• Modified MIT. Mistral Medium 3.5 (late April 2026), a 128B dense model that merges the former Magistral reasoning and Devstral 2 coding lines, publishes its weights under a “Modified MIT” licence: commercial use is permitted, with a carve-out for high-revenue companies. It is neither fully open nor API-only, so read the licence before you assume either.
• Commercial and premier. Codestral, Mistral OCR, Moderation 2 and the Embed models stay closed and are served only through Mistral’s hosted API.

New since October 2026: Mistral Large 4, in public preview, licence not yet named. On 6 October 2026 Mistral launched a public preview of Mistral Large 4 and promised open weights: “We will release the weights by the end of the month.” As of 25 October 2026 no Large 4 weights are on Mistral’s Hugging Face organisation and Mistral has not named the licence, so it belongs in none of the three tracks yet.

That gives you a spectrum: self-host an open model for full data control, or call a commercial model when you want a capability Mistral does not release as weights.

Current model lineup

The lineup turned over completely between December 2025 and April 2026. Mistral Large 2, Pixtral Large, Mistral 7B and the Mixtral models are retired from la Plateforme and appear only on Mistral’s deprecation list.

• Mistral Large 4 (mistral-large-4, version 26.10, public preview since 6 October 2026) is Mistral’s new top model and, in Mistral’s words, “our largest and most capable model to date”. The model page gives 1.05T total and 52B active parameters (the launch post rounds this to “a 1 trillion-parameter natively multimodal model”), a granular mixture-of-experts architecture, a 1.6B vision encoder and a 1M context window. It is a hybrid instruct-and-reasoning model with multimodal input, trained on 3,800 NVIDIA Grace Blackwell GPUs in Mistral’s own European data centres. Price on the model page, fetched 25 October 2026: $1.36 input, $0.14 cached input, $4.18 output per MTok. The 6 October changelog announced “Launch pricing: 50% off for 2 weeks”, which would have ended around 20 October; the page does not say whether $1.36 / $4.18 is the full rate or still the discounted one, so confirm on an invoice before budgeting. Weights are not published as of 25 October 2026 (Mistral promised them “by the end of the month”) and the licence is unnamed. Until then Mistral says it is red-teaming with “cybersecurity leaders, vetted partners, and state authorities, who will access the same model with reduced moderation and expanded cyber capabilities.” Vendor claims, not reproduced: 61.7% on DeepSWE v1.1, 28.3% on Terminal-Bench 4, 59.9% on AutomationBench, and 82% on one Artificial Analysis Cyber Index test, “the highest of any model”. On that last figure Mistral itself notes that Claude Opus 5.5 and GPT-6 Astra “score near zero on the same test because they refuse to perform the task”, so the lead measures willingness as much as capability.
• Mistral Large 3 (announced 2 December 2025) was the flagship until Large 4’s preview, and is still the largest Mistral model with published weights: a sparse mixture-of-experts model with roughly 675B total and 41B active parameters, a 256K context window, image input, and coverage of 40+ languages. Mistral released both the base and instruct weights under Apache 2.0.
• Mistral Medium 3.5 (late April 2026 - Mistral’s own sources vary between 28, 29 and 30 April) is a 128B dense model that folds the former Magistral reasoning line and Devstral 2 coding line into a single checkpoint with configurable reasoning effort, 256K context, and image input. Its weights are published under a “Modified MIT” licence that permits commercial use with a carve-out for high-revenue companies, so it is neither Apache 2.0 nor API-only.
• Mistral Small 4 (16 March 2026) is a 119B-total / ~6B-active MoE under Apache 2.0, with 256K context. It is the first Mistral model to unify reasoning, vision and agentic coding in one self-hostable checkpoint.
• Ministral 3 (2 December 2025) covers 14B, 8B and 3B, each shipping base, instruct and reasoning variants with image understanding, all Apache 2.0. These replace the retired Mistral 7B and Mixtral open-weight line.

Choosing between them is mostly a cost-versus-capability call: Ministral 3 or Small 4 for high-volume, cost-sensitive work, Large 3 for complex reasoning on weights you can hold today, Large 4 (preview) where you need 1M context or its stronger agentic and coding results and can accept a preview endpoint, and Medium 3.5 when you want reasoning effort you can tune per request.

Alongside its own catalogue, Mistral serves third-party Z.ai models on la Plateforme. GLM 5.3 (zai-glm-5-3) became generally available on 28 September 2026, and GLM 5.2 (zai-glm-5-2) is deprecated and retires on 31 October 2026, with GLM 5.3 “at the same price” as the replacement. On 22 October 2026 Mistral added GLM 5.3 Flash (zai-glm-5-3-flash), which the changelog describes as “A multimodal 320B-parameter (18B active) Mixture-of-Experts model from Z.ai with a 1M-token context window”, priced at $0.15 input, $0.03 cached input and $0.50 output per MTok.

Speech: Voxtral 3

On 20 October 2026 (changelog date; the announcement page carries 22 October in its metadata) Mistral released the Voxtral 3 family, all four models in Public Preview on the API: Voxtral TTS 3 (voxtral-mini-tts-3-0), Voxtral Mini Transcribe 3 (voxtral-mini-transcribe-3-0), Voxtral Mini Transcribe Realtime 3 (voxtral-mini-transcribe-realtime-3-0) and Voxtral Small Transcribe 3 (voxtral-small-transcribe-3-0), plus a managed Voice Agents pipeline and SDK. Mistral gives transcription in 50+ languages, TTS in “35+ languages and 50+ dialects”, Voxtral Mini Transcribe 3 at 4B parameters and $0.0035 per minute, and Voxtral TTS at $13 per 1M characters. The post says the family is available “for self-deployment on your own infrastructure”, but as of 25 October 2026 no Voxtral 3 weights are on Mistral’s Hugging Face organisation and no licence is named, so self-deployment appears to be by arrangement rather than by download. Mistral’s quality charts use gemini-3.8-flash as the judge for its TTS evaluations, which is worth knowing when reading them.

Time series: Timestral

Timestral (15 October 2026) is Mistral’s first family of time-series foundation models: four sizes, 5M, 28M, 123M and 209M parameters, producing probabilistic multivariate forecasts (21 quantiles) with past and known-future covariates, and running on CPU. The announcement lists the licence as “Modified MIT”, and the weights are on Hugging Face (mistralai/Timestral-209M and the smaller sizes, ungated, licence tagged “other”). Mistral says the 209M model “ranks first on all three of the field’s main public benchmarks, GIFT-Eval, TIME, and fev-bench”; it also says the results “have been submitted” to the leaderboards, so check the live leaderboards rather than the claim.

Specialist models: OCR and Lean

• Mistral OCR 4.1 (mistral-ocr-4-1) was released on 16 July 2026 and became generally available on 31 August 2026. The mistral-ocr-latest and mistral-ocr-4 aliases point to it. It is API-only and priced at $4 per 1,000 pages (see pricing below). If you pinned mistral-ocr-4-0 (June 2026), move to the 4.1 id or the alias.
• Leanstral 1.5 (labs-leanstral-1-5, 30 June 2026), the Apache 2.0 Lean 4 formal-proof engineering model, retired from the API on 30 September 2026 (as did OCR 4.0, mistral-ocr-4-0); Mistral announced the date when it released the model. The models page lists no hosted successor, so if you rely on it, plan to self-host the open weights or switch provider before then. (Its predecessor, labs-leanstral-2603, retired on 30 June 2026.)

Company news: Pimento acquisition

Sifted reported on 22 September 2026, citing company filings, that Mistral is acquiring Pimento, a Paris-based startup that builds AI-assisted advertising creation, in a cash-and-shares deal worth €12.7 million. Sifted and FW.media both describe it as Mistral’s third acquisition of 2026, after Koyeb (infrastructure) and Emmi AI (industrial engineering). FW.media and other coverage link the purchase to Mistral’s Vibe assistant and business applications. Mistral had not published its own announcement in its changelog as of 25 September 2026, so treat the deal terms as reported rather than confirmed.

Where Mistral sits in your stack

Mistral is a model provider, not a full application platform. It supplies the intelligence layer that your application calls, whether you self-host the weights or hit the hosted API.

How to access it

You reach Mistral three ways, depending on how much control you want.

Le Chat, now Vibe. The consumer-facing chat product, comparable to other chat assistants. It runs Mistral’s models behind a web and mobile interface. Mistral rebranded it to Vibe in May 2026; sources disagree on the exact day, and the Vibe name had already been used for a separate terminal coding agent earlier that year. Use it to try the models before building anything.

The hosted API. Mistral serves both open-weight and commercial models through la Plateforme, an API with a developer console. You send a prompt, you get a completion, and Mistral runs the inference on its own infrastructure. Mistral states that its servers are hosted in the EU, which matters for teams with data-residency requirements. The platform is no longer Mistral-only: it also resells a third-party model, Z.ai’s GLM 5.2 with a 1M context window, which is currently the longest context you can reach through a Mistral endpoint.

Self-hosting the open weights. For the open-weight models, you download the weights and run them on your own GPUs or through a third-party inference host. This keeps every request inside your own perimeter. It costs more operationally, and you own the scaling and reliability work.

Using the API

Mistral’s API is OpenAI-compatible. Use the official SDK, or point the openai package at Mistral’s base URL.

pip install mistralai

from mistralai import Mistral

client = Mistral(api_key="YOUR_MISTRAL_API_KEY")

response = client.chat.complete(
    model="mistral-large-latest",
    messages=[{"role": "user", "content": "Summarise the EU AI Act in three bullet points."}]
)
print(response.choices[0].message.content)

Via the openai SDK (drop-in replacement):

from openai import OpenAI

client = OpenAI(
    api_key="YOUR_MISTRAL_API_KEY",
    base_url="https://api.mistral.ai/v1"
)

chat = client.chat.completions.create(
    model="mistral-small-latest",
    messages=[{"role": "user", "content": "What is RAG?"}]
)

Function calling

Mistral supports OpenAI-compatible function calling on all large models.

import json
from mistralai import Mistral

client = Mistral(api_key="YOUR_MISTRAL_API_KEY")

tools = [
    {
        "type": "function",
        "function": {
            "name": "get_company_info",
            "description": "Return firmenbuch (company register) data for an Austrian company.",
            "parameters": {
                "type": "object",
                "properties": {
                    "company_name": {"type": "string", "description": "Legal name of the company"},
                    "country": {"type": "string", "enum": ["AT", "DE", "CH"]}
                },
                "required": ["company_name", "country"]
            }
        }
    }
]

response = client.chat.complete(
    model="mistral-large-latest",
    messages=[{"role": "user", "content": "Look up Erste Bank AG in Austria."}],
    tools=tools,
    tool_choice="auto"
)

tool_call = response.choices[0].message.tool_calls[0]
args = json.loads(tool_call.function.arguments)
print(args)  # {'company_name': 'Erste Bank AG', 'country': 'AT'}

Codestral for code generation

Codestral remains Mistral’s dedicated code-completion model, currently at version stamp v25.08. Unlike the general-purpose lineup it sits in the Premier tier with closed weights, so you call it rather than self-host it. The older MNPL-licensed 22B Codestral is superseded. If you want a self-hostable coding model, Mistral now points you at Small 4 or Medium 3.5, which absorbed the Devstral agentic-coding line. Check the models page for the current context window before you design around it.

client = Mistral(api_key="YOUR_MISTRAL_API_KEY")

response = client.chat.complete(
    model="codestral-latest",
    messages=[
        {
            "role": "user",
            "content": "Write a FastAPI endpoint that accepts a PDF and returns extracted text using AWS Textract."
        }
    ]
)
print(response.choices[0].message.content)

Pricing (la Plateforme, as of 9 September 2026; Large 4 and GLM rows added 25 October 2026)

Mistral publishes per-model rates in US dollars on its API pricing page. Earlier versions of this page quoted euro figures for a lineup that no longer exists.

The Large 4 row is the model page’s figure on 25 October 2026 (cached input $0.14), with the launch-discount caveat above; the other Mistral rows were not re-checked on 25 October. The specialised models price on their own units: Mistral OCR 4.1 at $4 per 1,000 pages, Voxtral TTS at $0.016 per 1,000 characters, and Voxtral Mini Transcribe at $0.003 per audio minute. For Voxtral 3, the announcement gives $0.0035 per minute for Voxtral Mini Transcribe 3 and $13 per 1M characters for Voxtral TTS 3.

Typical use

Teams reach for Mistral when European data residency or the option to self-host is a hard requirement, not a nice-to-have. Common patterns:

• Regulated workloads where data cannot leave EU infrastructure, so an EU-hosted API or self-hosted weights is the deciding factor.
• On-premise or private-cloud deployment using an Apache 2.0 open-weight model, where owning the weights removes vendor lock-in.
• Cost-sensitive backends that run a smaller open model locally instead of paying per-token for a commercial API.
• Multilingual and code tasks. Mistral Large 3 covers 40+ languages, and Mistral ships dedicated models for coding (Codestral), speech (Voxtral), document OCR, embeddings and moderation alongside the general-purpose line.

How it compares

At the provider level, the distinguishing axes are where the company is based, whether you can get the weights, and where inference runs.

Model against model, the flagship comparison looks like this:

Only the Mistral column is quoted from Mistral’s own published price list. GPT-5.6 and Llama figures above are indicative - check OpenAI API for current context windows and pricing across its Sol/Terra/Luna tiers. Claude Sonnet 5’s $2.00/MTok input price is now Anthropic’s permanent standard rate (see Claude Anthropic for the full current lineup and pricing table). Note that Llama is no longer Meta’s frontier line - see Meta Llama for what replaced it. The 2026 LLM landscape comparison places all these providers side by side.

When not to use it

• You want a single API across many vendors. A managed aggregator like Amazon Bedrock or Azure OpenAI lets you switch models without changing providers.
• You need a million-token context on a generally available model. Mistral’s own GA models top out at 256K; Large 4 has 1M but is a public preview. For 1M context, use Gemini or Claude Opus 5.5 , or the third-party GLM 5.3 and GLM 5.3 Flash that Mistral serves on la Plateforme, which carry a 1M window.
• You need video or audio-native reasoning. Large 3, Medium 3.5, Small 4 and Ministral 3 all accept image input, so still images are no longer a reason to route elsewhere. Speech is handled by the separate Voxtral models rather than in the chat models, and there is no video input. For a single model that reasons over video, use Gemini .
• You need the strongest available reasoning right now. Benchmark the specific task against Claude and others rather than assuming any single provider leads. See how AI models are evaluated .
• You have no data-residency or self-host requirement. Mistral’s main differentiators are EU hosting and open weights. Without those needs, choose on capability and price alone.
• You lack the operations capacity to self-host. Running open weights yourself means owning GPU provisioning, scaling, and uptime. If you cannot staff that, stay on a hosted API.

Further reading

• What is an LLM? : plain-English explanation of large language models.
• What are foundation models? : the broader model category Mistral builds within.
• The 2026 LLM landscape : how providers compare across capability and access.
• Alibaba Qwen : another provider that ships open-weight models.
• Building RAG systems : how to build a retrieval-augmented generation pipeline with any LLM provider.
• LLM gateway architecture : route between Mistral, OpenAI, and Anthropic with fallback logic.
• la Plateforme documentation : API reference, model cards, rate limits.
• Mistral open-weight models : which models ship as downloadable weights, and under which licence.

Sources

• Mistral AI homepage
• Mistral models and deprecations : the authoritative current lineup, version stamps, and retired models.
• Mistral changelog (fetched 25 October 2026): Mistral Large 4 public preview and the 50%-off launch pricing (6 October 2026), Voxtral 3 (20 October), GLM 5.3 Flash and its price (22 October), GLM 5.3 GA (28 September) and the GLM 5.2 retirement on 31 October 2026, OCR 4.1 released 16 July and GA 31 August 2026, Leanstral 1.5 and OCR 4.0 retired 30 September 2026.
• Introducing Mistral Large 4, Mistral, 6 October 2026 : parameters, training hardware, the weights promise, the red-teaming access, and the benchmark claims including the refusal note.
• Mistral Large 4 model page (fetched 25 October 2026): version 26.10, Public Preview, 1.05T / 52B, 1.6B vision encoder, 1M context, $1.36 / $0.14 / $4.18.
• Voxtral 3, Mistral, October 2026 : models, languages, prices and the self-deployment statement.
• Timestral, Mistral, 15 October 2026 : sizes, benchmarks and the Modified MIT licence; weights at huggingface.co/collections/mistralai/timestral .
• Exclusive: Mistral acquires adtech startup Pimento in €12.7m cash-and-shares deal, filings show, Sifted, 22 September 2026 (secondary source)
• MISTRAL AI acquires Pimento, moving further into business applications, FW.media, September 2026 (secondary source)
• Mistral open-weight model list and licences
• Mistral 3 announcement, 2 December 2025 : Mistral Large 3 and Ministral 3 under Apache 2.0.
• Mistral Small 4 announcement, 16 March 2026
• Mistral Medium 3.5 model card : the Modified MIT licence terms.
• Mistral API pricing : current per-model USD rates.
• Mistralai Python SDK on PyPI : package source and changelog.
• Mistral AI on Wikipedia