Mistral technology

AI models

We release the world’s most capable open models, enabling frontier AI innovation.

Developer platform

Our portable developer platform serves our open and optimized models for building fast and intelligent applications. We offer flexible access options!

Open source models

We’re committed to empower the AI community with open technology. Our open models sets the bar for efficiency, and are available for free under Apache 2.0, a fully permissive license, that allows to use the models anywhere without any restriction.

Mistral 7B

Our very first. A 7B transformer model, fast-deployed and easily customisable. Small, yet very powerful for a variety of use cases.

  • Performant in English and code
  • 32k context window
Mixtral 8x7B

A 7B sparse Mixture-of-Experts (SMoE). Uses 12.9B active parameters out of 45B total.

  • Fluent in English, French, Italian, German, Spanish, and strong in code
  • 32k context window
Mixtral 8x22B

Mixtral 8x22B is currently the most performant open model. A 22B sparse Mixture-of-Experts (SMoE). Uses only 39B active parameters out of 141B.

  • Fluent in English, French, Italian, German, Spanish, and strong in code
  • 64k context window
  • Native function calling capacities
  • Function calling and json mode available on our API endpoint

Optimized models

Our optimized commercial models are designed for performance and are available via our flexible deployment options.

Mistral Small

Cost-efficient reasoning for low-latency workloads.

  • Fluent in English, French, Italian, German, Spanish, and strong in code
  • Context window of 32k tokens, with excellent recall for retrieval augmentation
  • Native function calling capacities, JSON outputs
  • Concise, useful, unopinionated, and fully modular
Mistral Large

Top-tier reasoning for high-complexity tasks. The most powerful model of the Mistral AI family.

  • Fluent in English, French, Italian, German, Spanish, and strong in code
  • Context window of 32k tokens, with excellent recall for retrieval augmentation
  • Native function calling capacities, JSON outputs
  • Concise, useful, unopinionated, and fully modular
Mistral Embed

State-of-the-art semantic for extracting representation of text extracts.

  • English only for now
  • Achieves a retrieval score of 55.26 on the Massive Text Embedding Benchmark (MTEB)

Codestral

Codestral is an open-weight generative AI model explicitly designed for code generation tasks. It helps developers write and interact with code through a shared instruction and completion API endpoint. As it masters code and English, it can be used to design advanced AI applications for software developers.

Codestral 22B

State-of-the-art Mistral model trained specifically for code tasks.

Mistral codestral 22B
  • Trained on 80+ programming languages (incl. Python, Java, C, C++, PHP, Bash)
  • Optimized for low latency: Way smaller than competitive coding models
  • Base version (for code completion and generation) via La Plateforme as well as through our partners and Instruct version via Le Chat
  • Context window of 32k
  • Available under Mistral AI Non-Production License
Mistral codestral 22B

Pareto-optimal Apache models

We’re constantly innovating to provide the most capable and efficient models.

Cost vs Performance
Cost vs Performance

We build models that are highly efficient

We build models that offer unparalleled cost efficiency for their respective sizes, delivering the best performance-to-cost ratio available on the market. Mixtral 8x22B is the most powerful open source model with significantly fewer parameters than its competition:

  • Much faster and performant than any 70B model
  • Outperforms Command R+ by being more than 2.5 times smaller
  • Mixtral 8x7B outperforms Llama 2 70B on most benchmarks with 6x faster inference

Under Apache 2.0

Our open models are truly open source, licensed under Apache 2.0, a fully permissive license that allows for unrestricted use in any context.

La Plateforme

Access our latest products via our developer platform, hosted in Europe

from mistralai.client import MistralClient
from mistralai.models.chat_completion import ChatMessage

api_key = os.environ["MISTRAL_API_KEY"]
model = "mistral-tiny"

client = MistralClient(api_key=api_key)

messages = [
    ChatMessage(role="user",
    content="Who is the most renowned French painter?")
]
Built for developers

La Plateforme is developers’ preferred way to access all Mistral Al’s models. Hosted and served on Mistral Al infrastructure, in Europe.

  • Our best models at the best price : Get access to our models at an unmatched price/performance point
  • Guides & community : Use our guides and community forums to build your own application and services
  • Secure by design : Your data are encrypted at rest (AES256) and in transit (TLS 1.2+); our servers are in the EU
from mistralai.client import MistralClient
from mistralai.models.chat_completion import ChatMessage

api_key = os.environ["MISTRAL_API_KEY"]
model = "mistral-tiny"

client = MistralClient(api_key=api_key)

messages = [
    ChatMessage(role="user",
    content="Who is the most renowned French painter?")
]

Fine-tune our models

We allow you to fine-tune our models in an easy, effective & cost-efficient way, and thus use smaller and better-suited models to solve your specific use cases. Fine-tuning can be done with our open-source fine-tuning code as well as on La Plateforme with our efficient Fine-tuning API.

Use Mistral fine-tuning code

Benefit from Mistral fine-tuning code to perform fine-tuning on Mistral open-source models on your own.

Fine-tune Mistral models on La Plateforme with Mistral Fine-tuning API

Leverage Mistral’s unique expertise in training models by using our highly efficient fine-tuning service to specialize both our open-source and commercial models.

Deploy anywhere

La Plateforme
La Plateforme

Get started with Mistral models in a few clicks via our developer platform hosted on Mistral’s infrastructure and build your own applications and services. Our servers are hosted in EU.

Cloud platforms
Cloud platforms

Access our models via your preferred cloud provider and use your cloud credits. Our open models are currently available via our cloud partners (GCP, AWS, Azure, Snowflake, NVIDIA).
Mistral Large is available on Azure AI Studio, AWS Bedrock and Snowflake.

Self-deployment
Self-deployment

Deploy Mistral models on virtual cloud or on-prem. Self-deployment offers more advanced levels of customisation and control. Your data stays within your walls. Try deploying our open models, and contact our team to deploy our optimized models similarly.

Pay-as-you-go pricing

Price in $
Price in €
Open source models
InputOutput
open-mistral-nemo-2407Mistral NeMo is a state-of-the-art 12B model developed with NVIDIA.$0.3 /1M tokens$0.3 /1M tokens
open-mistral-7bA 7B transformer model, fast-deployed and easily customisable.$0.25 /1M tokens$0.25 /1M tokens
codestral-mamba-latestA Mamba 2 language model specialized in code generation.$0.25 /1M tokens$0.25 /1M tokens
open-mixtral-8x7bA 7B sparse Mixture-of-Experts (SMoE). Uses 12.9B active parameters out of 45B total.$0.7 /1M tokens$0.7 /1M tokens
open-mixtral-8x22bMixtral 8x22B is currently the most performant open model. A 22B sparse Mixture-of-Experts (SMoE). Uses only 39B active parameters out of 141B.$2 /1M tokens$6 /1M tokens
InputOutput
open-mistral-nemo-2407Mistral NeMo is a state-of-the-art 12B model developed with NVIDIA.0.27€ /1M tokens0.27€ /1M tokens
open-mistral-7bA 7B transformer model, fast-deployed and easily customisable.0.2€ /1M tokens0.2€ /1M tokens
codestral-mamba-latestA Mamba 2 language model specialized in code generation.0.2€ /1M tokens0.2€ /1M tokens
open-mixtral-8x7bA 7B sparse Mixture-of-Experts (SMoE). Uses 12.9B active parameters out of 45B total.0.65€ /1M tokens0.65€ /1M tokens
open-mixtral-8x22bMixtral 8x22B is currently the most performant open model. A 22B sparse Mixture-of-Experts (SMoE). Uses only 39B active parameters out of 141B.1.9€ /1M tokens5.6€ /1M tokens
Optimized models
InputOutput
mistral-small-2402Cost-efficient reasoning for low-latency workloads.$1 /1M tokens$3 /1M tokens
codestral-2405State-of-the-art Mistral model trained specifically for code tasks.$1 /1M tokens$3 /1M tokens
mistral-medium-2312*Will soon be deprecated*$2.7 /1M tokens$8.1 /1M tokens
mistral-large-2402Top-tier reasoning for high-complexity tasks. The most powerful model of the Mistral AI family.$4 /1M tokens$12 /1M tokens
InputOutput
mistral-small-2402Cost-efficient reasoning for low-latency workloads.0.9€ /1M tokens2.8€ /1M tokens
codestral-2405State-of-the-art Mistral model trained specifically for code tasks.0.9€ /1M tokens2.8€ /1M tokens
mistral-medium-2312*Will soon be deprecated*2.5€ /1M tokens7.5€ /1M tokens
mistral-large-2402Top-tier reasoning for high-complexity tasks. The most powerful model of the Mistral AI family.3.8€ /1M tokens11.3€ /1M tokens
Embeddings
InputOutput
mistral-embedState-of-the-art semantic for extracting representation of text extracts.$0.1 /1M tokens$0.1 /1M tokens
InputOutput
mistral-embedState-of-the-art semantic for extracting representation of text extracts.0.1€ /1M tokens$0.1 /1M tokens
Fine-tuning models
One-off trainingStorageInputOutput
Mistral 7B$2 /1M tokens$2 per month per model$0.75 /1M tokens$0.75 /1M tokens
Mistral Small$4 /1M tokens$2 per month per model$2.5 /1M tokens$7.5 /1M tokens
One-off trainingStorageInputOutput
Mistral 7B1.9€ /1M tokens1.9€ per month per model0.7€ /1M tokens0.7€ /1M tokens
Mistral Small3.8€ /1M tokens1.9€ per month per model2.3€ /1M tokens7€ /1M tokens
Fine-tuning Pricing Explained

Mistral AI provides a fine-tuning API through La Plateforme, making it easy to fine-tune our open-source and commercial models. There are three costs related to fine-tuning:

  • One-off training: Price per token on the data you want to fine-tune our standard models on; minimum fee per fine-tuning job of $4
  • Inference: Price per input/output token when using the fine-tuned model(s)
  • Storage: Price per month per model for storage (irrespective of model usage; models can be deleted any time)