Mistral AI

Models and offerings #

Who they are

This hub is for people who want an open-weight model with a licence their legal team will accept, and for European teams who need a vendor under EU jurisdiction. Mistral AI was founded in Paris in April 2023 by researchers from Google DeepMind and Meta, and released Mistral 7B that September under Apache-2.0, followed by Mixtral 8x7B in December, the model that made sparse mixture-of-experts a practical open architecture. It has since raised several billion euros and sells to governments and large European companies.

Its position in the field is the permissive-licence counterweight to Meta and Google: most of its small and mid-size releases are Apache-2.0 with no use restrictions, while its largest models and its speech and OCR products are either API-only (“Premier” in Mistral’s catalogue) or under a custom open-weight licence.

Tooling and APIs

Keys come from the Studio console at console.mistral.ai (the platform was called La Plateforme until 2025 and older documentation still uses that name). pip install mistralai installs the SDK, and the API also accepts OpenAI-format requests. Open-weight models are on Hugging Face under mistralai and need mistral-common for the tokeniser, and vLLM, llama.cpp and Ollama all support them. Mistral models also appear on Azure, AWS Bedrock and Google Vertex AI.

# needs-api
import os
from mistralai import Mistral

client = Mistral(api_key=os.environ["MISTRAL_API_KEY"])
response = client.chat.complete(
    model="mistral-large-4",
    messages=[{"role": "user", "content": "In one sentence, what does a learning rate control?"}],
)
print(response.choices[0].message.content)

The current models, per Mistral’s documentation: Mistral Large 4 (about 1.05 T total parameters, 52 B active according to the docs and 49 B according to the launch post, 1 M context, API preview from 6 October 2026, weights promised for late October and listed in the docs under the Mistral Ecosystem Open Weight License, although the launch post’s model card said Apache-2.0). Mistral Small 4 (119 B total, 6.5 B active, 256 k context, Apache-2.0, 16 March 2026) replaces the earlier Small 3.x, Magistral and Devstral Small lines in one model. Ministral 3 (3 B, 8 B and 14 B, Apache-2.0, December 2025) covers the edge. Codestral, Codestral Embed, Mistral Embed, OCR 4.1, the Voxtral transcription and TTS models and Mistral Moderation 2 are API-only, except Voxtral Small whose weights are Apache-2.0.

Pricing at the time of writing: the Large 4 launch post gives 0.70 USD per million input tokens, 0.07 USD cached and 2.10 USD output, while the general pricing page lists “Mistral Large” at 0.50 and 1.50, so check mistral.ai/pricing for the current table. Batch is half price. Vibe Team is 24.99 USD per user a month. Pain points: model names carry two-digit year-month versions (mistral-small-2603) that make migration grep-able but confusing, the docs and blog disagree on numbers, and the retirement calendar is aggressive (below).

Migration notes

From OpenAI: the mistralai SDK mirrors the OpenAI shape closely (chat.complete, messages, tools), and the API’s OpenAI-compatible mode means base_url plus a model id is often enough. Function calling and JSON mode exist but reasoning controls differ.

Open weights to and from Llama: the architecture is Llama-like, so adapters and serving stacks transfer, but the tokeniser (Tekken, a tiktoken-style tokeniser on recent models) and the chat template are Mistral’s own. Use mistral-common rather than hand-rolled prompts. GGUF conversions are published by the community within days of a release, and NVFP4 and FP8 variants of Small 4 come from Mistral itself.

To self-hosting: Small 4 at 119 B total parameters wants two 80 GB GPUs in BF16 or one with FP8 quantisation, while Ministral 3 14 B runs on a laptop. Large 4 is a multi-node model. The quantisation formats reference covers the trade-offs.

Watch for

Deprecations. Mistral’s model overview lists a retirement date for nearly every older id: Mistral Medium 3.5 and Large 3 are deprecated from 31 October 2026 and retired on 30 April 2027, and earlier Small, Magistral and Devstral ids retire on a rolling six-month schedule. If you pin a dated model you must plan the next move when you do it.

Licences vary per model and have changed over time. Apache-2.0 means no restrictions. The Mistral Research License and the Non-Production License on some 2024 and 2025 checkpoints forbid commercial use. The Mistral Ecosystem Open Weight License on Large 4 is a modified MIT licence whose terms, according to third-party reports we could not verify against the full text, exclude third-party hosting providers, so read it before serving Large 4 as a product. Open weights under a custom licence are open weights, restricted.

Benchmark claims in Mistral’s launch posts (for example 61.7 percent on DeepSWE 1.1 for Large 4) are the vendor’s reports. The company’s own figures for active parameters differ between its blog and its documentation, which is a useful reminder to read the model card rather than the headline.