Skip to main content

Models

Overview​

Mistral provides two types of models: free models and premier models.

Premier models​

ModelWeight availabilityAvailable via APIDescriptionMax TokensAPI EndpointsVersion
Mistral Large✔️
Mistral Research License		✔️Our top-tier reasoning model for high-complexity tasks with the lastest version v2 released July 2024. Learn more on our blog post128kmistral-large-latest24.07
Mistral Small✔️
Mistral Research License		✔️Our latest enterprise-grade small model with the lastest version v2 released September 2024. Learn more on our blog post128kmistral-small-latest24.09
Codestral✔️
Mistral Non-Production License		✔️Our cutting-edge language model for coding released May 202432kcodestral-latest24.05
Mistral Embed✔️Our state-of-the-art semantic for extracting representation of text extracts8kmistral-embed23.12

Free models​

  • Latest models
ModelWeight availabilityAvailable via APIDescriptionMax TokensAPI EndpointsVersion
Pixtral✔️
Apache2		✔️A 12B model with image understanding capabilities in addition to text. Learn more on our blog post128kpixtral-12b-240924.09
  • Research models
ModelWeight availabilityAvailable via APIDescriptionMax TokensAPI EndpointsVersion
Mistral Nemo✔️
Apache2		✔️Our best multilingual open source model released July 2024. Learn more on our blog post128kopen-mistral-nemo24.07
Codestral Mamba✔️
Apache2		✔️Our first mamba 2 open source model released July 2024. Learn more on our blog post256kopen-codestral-mambav0.1
Mathstral✔️
Apache2		Our first math open source model released July 2024. Learn more on our blog post32kNAv0.1
  • Legacy models
ModelWeight availabilityAvailable via APIDescriptionMax TokensAPI EndpointsVersion
Mistral 7B✔️
Apache2		✔️Our best open source model to date released April 2024. Learn more on our blog post32kopen-mistral-7bv0.3
Mixtral 8x7B✔️
Apache2		✔️Our first sparse mixture-of-experts released December 2023. Learn more on our blog post32kopen-mixtral-8x7bv0.1
Mixtral 8x22B✔️
Apache2		✔️Our first dense model released September 2023. Learn more on our blog post64kopen-mixtral-8x22bv0.1

API versioning​

Mistral AI API are versions with specific release dates. To prevent any disruptions due to model updates and breaking changes, it is recommended to use the dated versions of the Mistral AI API. Additionally, be prepared for the deprecation of certain endpoints in the coming months.

Here are the details of the available versions:

  • open-mistral-nemo: currently points to open-mistral-nemo-2407.
  • mistral-small-latest: currently points to mistral-small-2409. mistral-small-2402 is deprecated.
  • mistral-medium-latest: currently points to mistral-medium-2312. The previous mistral-medium has been dated and tagged as mistral-medium-2312. Mistral Medium will be deprecated shortly.
  • mistral-large-latest: currently points to mistral-large-2407. mistral-large-2402 will be deprecated shortly.
  • codestral-latest: currently points to codestral-2405.

Benchmarks results​

Mistral ranks second among all models generally available through an API. It offers top-tier reasoning capabilities and excels in multilingual tasks and code generation.

You can find the benchmark results in the following blog posts:

  • Mistral 7B: outperforms Llama 2 13B on all benchmarks and Llama 1 34B on many benchmarks.
  • Mixtral 8x7B: outperforms Llama 2 70B on most benchmarks with 6x faster inference and matches or outperforms GPT3.5 on most standard benchmarks. It handles English, French, Italian, German and Spanish, and shows strong performance in code generation.
  • Mixtral 8x22B: our most performant open model. It handles English, French, Italian, German, Spanish and performs strongly on code-related tasks. Natively handles function calling.
  • Mistral Large: a cutting-edge text generation model with top-tier reasoning capabilities. It can be used for complex multilingual reasoning tasks, including text understanding, transformation, and code generation.
  • Codestral: as a 22B model, Codestral sets a new standard on the performance/latency space for code generation compared to previous models used for coding.
  • Codestral-Mamba: we have trained this model with advanced code and reasoning capabilities, enabling the model to have a strong performance on par with SOTA transformer-based models.
  • Mathstral: Mathstral stands on the shoulders of Mistral 7B and specialises in STEM subjects. It achieves state-of-the-art reasoning capacities in its size category across various industry-standard benchmarks.
  • Mistral Nemo: Mistral Nemo's reasoning, world knowledge, and coding performance are state-of-the-art in its size category. As it relies on standard architecture, Mistral Nemo is easy to use and a drop-in replacement in any system using Mistral 7B that it supersedes.
  • Pixtral: Pixtral 12B the first open-source model to demonstrate state-of-the-art multimodal understanding, without regressing on abilities in pure text.

Picking a model​

This guide will explore the performance and cost trade-offs, and discuss how to select the appropriate model for different use cases. We will delve into various factors to consider, offering guidance on choosing the right model for your specific needs.

Today, Mistral models are behind many LLM applications at scale. Here is a brief overview on the types of use cases we see along with their respective Mistral model:

  1. Simple tasks that one can do in bulk (Classification, Customer Support, or Text Generation) are powered by Mistral Small.
  2. Intermediate tasks that require moderate reasoning (Data extraction, Summarizing a Document, Writing emails, Writing a Job Description, or Writing Product Descriptions) are powered by Mistral 8x22B.
  3. Complex tasks that require large reasoning capabilities or are highly specialized (Synthetic Text Generation, Code Generation, RAG, or Agents) are powered by Mistral Large.

Performance and cost trade-offs​

When selecting a model, it is essential to evaluate the performance, and cost trade-offs. Depending on what’s most important for your application, your choice may differ significantly. Note that the models will be updated over time, the information we share below only reflect the current state of the models.

Notably, Mistral Large v2 is currently outperforming all of our other models across almost all benchmarks. By considering the performance, speed, and cost details, hopefully you can find the best model that suits your application's needs.

As a general rule, if you are new to using Large Language Models, you can always start with Mistral Large first. This will allow you to evaluate its cost and performance and determine whether you need to downgrade to Mistral Small or even Mistral Nemo. If you are already using another large language model and are looking to replace a particularly powerful one, then Mistral Large may be the best choice for you.

Use cases​

Mistral Nemo: Simple tasks that one can do in bulk​

Mistral Nemo is the ideal choice for simpe tasks that one can do in builk - like Classification, Customer Support, or Text Generation. It offers excellent performance at an affordable price point. For instance, it can be effectively used for a classification task to classify if an email is spam or not:

Prompt:

Classify the following email to determine if it is spam or not. Only respond with the exact words "Spam" or "Not spam".

🎉 Urgent! You've Won a $1,000,000 Cash Prize! 💰 To claim your prize, please click on the link below: https://bit.ly/claim-your-prize

All of our models can accurately classify this email correctly as “Spam”. Mistral Nemo is capable to provide the correct classification as the larger models. So it is the most efficient and affordable choice for this kind of tasks.

Mistral Small: Intermediate tasks that require language transformation​

Mistral Small is the ideal for intermediate tasks that require moderate reasoning - like Data extraction, Summarizing a Document, Writing a Job Description, or Writing Product Descriptions. Mistral Small strikes a balance between performance and capability, making it suitable for a wide range of tasks that only require language transformaion. For example, Mistral Small can write an email:

Prompt:

Compose a welcome email for new customers who have just made their first purchase with your product. 
Start by expressing your gratitude for their business, and then convey your excitement for having them as a customer. 
Include relevant details about their recent order. Sign the email with "The Fun Shop Team".

Order details:
- Customer name: Anna
- Product: hat 
- Estimate date of delivery: Feb. 25, 2024
- Return policy: 30 days

Output:

Subject: Welcome to The Fun Shop, Anna! Your Hat is on its Way!

Dear Anna,

We hope this message finds you well. We are thrilled to welcome you to The Fun Shop family! We are incredibly grateful for your business and are excited to have you as our new customer.

We are processing your recent order for the fabulous hat you've selected. Your new accessory is set to be delivered by Feb. 25, 2024. We hope it arrives just in time to add a touch of fun and style to your outfits!

Please remember that our return policy allows you to return any item within 30 days of receipt if you are not completely satisfied. We want you to love every item you purchase from us, so if there's any issue, please don't hesitate to reach out.

Thank you once again for choosing The Fun Shop. We look forward to serving you and making your shopping experience a fun and enjoyable one!

Best Regards,
The Fun Shop Team

Mistral Large: Complex tasks that require large reasoning capabilities or are highly specialized​

Mistral Large is ideal for complex tasks that require large reasoning capabilities or are highly specialized - like Synthetic Text Generation, Code Generation, RAG, or Agents.

  • Complex reasoning

Mistral Large outperforms our other four models in commonsense and reasoning benchmarks, making it the best choice for complex reasoning tasks.

In this example, we ask Mistral Large to calculate the difference in payment dates between the two customers whose payment amounts are closest to each other in the given dataset. Mistral Large can easily and automatically break down the problem into smaller steps: identify the two closest payment amounts, find their respective payment dates, and calculate the date difference. Mistral Large is an excellent choice for tasks that require complex reasoning and problem-solving skills.

Prompt:

Calculate the difference in payment dates between the two customers whose payment amounts are closest to each other in the given dataset: 

'{
  "transaction_id":{"0":"T1001","1":"T1002","2":"T1003","3":"T1004","4":"T1005"},
    "customer_id":{"0":"C001","1":"C002","2":"C003","3":"C002","4":"C001"},
    "payment_amount":{"0":125.5,"1":89.99,"2":120.0,"3":54.3,"4":210.2},
"payment_date":{"0":"2021-10-05","1":"2021-10-06","2":"2021-10-07","3":"2021-10-05","4":"2021-10-08"},
    "payment_status":{"0":"Paid","1":"Unpaid","2":"Paid","3":"Paid","4":"Pending"}
}'

Output:

To find the difference in payment dates between the two customers whose payment amounts are closest to each other, we first need to identify the two closest payment amounts and then find their respective payment dates.

The payment amounts are: 125.5, 89.99, 120.0, 54.3, 210.2

The two closest payment amounts are 125.5 and 120.0. The payment date for the payment amount 125.5 is '2021-10-05' and for the payment amount 120.0 is '2021-10-07'.

The difference in payment dates can be calculated as follows:

Date difference = |'2021-10-07' - '2021-10-05'| = 2 days

So, the difference in payment dates between the two customers whose payment amounts are closest to each other is 2 days.

Let’s try another prompt we found online that requires reasoning and logic understanding. Mistral Large is able to answer it correctly.

Prompt:

There's an apple inside a blue box. There's also a red box inside the blue box. The red box has a lid on it. How can I get the apple?

Output:

You can get the apple by opening the blue box. The apple is not inside the red box, so you don't need to open the red box. Simply reach into the blue box and take the apple.

Here is another example. Mistral Large is able to answer it correctly.

Prompt:

Which one is heavier a pound of iron or a kilogram of feather

Output:

A kilogram of feathers is heavier than a pound of iron.

A pound is a unit of weight that is equivalent to approximately 0.453592 kilograms. Therefore, a pound of iron weighs less than a kilogram of feathers.

It may seem counterintuitive because we often think of iron as being a heavy material and feathers as being light, but the weight of an object depends on its mass and the force of gravity acting upon it, not on its material or composition.

  • Coding

Mistral Large, the top performer in coding tasks, is the ideal choice for users who priortize coding capabilities in their model selection.

Here is an example of Mistral Large giving the correct code following the prompt.

Prompt:

Write a function to find the maximum number of segments of lengths a, b and c that can be formed from n.

Output:

# Here is a possible function in Python to find the maximum number of segments that can be formed from a given length `n` using segments of lengths `a`, `b`, and `c`:

def max_segments(n, a, b, c):
    # Initialize the maximum number of segments to 0
    max_num_segments = 0

    # Loop through all possible combinations of segments
    for i in range(n // a + 1):
        for j in range(n // b + 1):
            for k in range(n // c + 1):
                # Check if the combination is valid and update the maximum number of segments
                if i * a + j * b + k * c == n:
                    max_num_segments = max(max_num_segments, i + j + k)

    return max_num_segments

# This function uses nested loops to generate all possible combinations of segments of lengths `a`, `b`, and `c`, respectively. For each combination, it checks if the total length of the segments is equal to `n`, and if so, it updates the maximum number of segments found so far. The function returns the maximum number of segments that can be formed from `n`.

Here is another example of Mistral Large writing a function for computing square roots using the babylonian method.

Prompt:

You are an expert Python programmer, and here is your task: Write a function for computing square roots using the babylonian method. Your code should pass these tests:

assert math.isclose(babylonian_squareroot(10), 3.162277660168379, rel_tol=0.001)
assert math.isclose(babylonian_squareroot(2), 1.414213562373095, rel_tol=0.001)
assert math.isclose(babylonian_squareroot(9), 3.0, rel_tol=0.001)

Output:

def babylonian_squareroot(number):
    if(number == 0):
        return 0;
    g = number/2.0;
    g2 = g + 1;
    while(g != g2):
        n = number/ g;
        g2 = g;
        g = (g + n)/2;
    return g;"
  • Multi-lingual tasks

In addition to its exceptional performance in complex reasoning tasks and coding tasks, Mistral Large also demonstrates superior capabilities in handling multi-lingual tasks. Mistral-large has been specifically trained to understand and generate text in multiple languages, especially in French, German, Spanish and Italian. Mistral Large can be especially valuable for businesses and users that need to communicate in multiple languages.

Prompt:

Lequel est le plus lourd une livre de fer ou un kilogramme de plume

Output:

Un kilogramme de plumes est plus lourd qu'une livre de fer.

Pour comprendre cela, il est important de se rappeler que :

1. Une livre (lb) est une unité de mesure de masse principalement utilisée aux États-Unis et dans d'autres systèmes de mesure impériaux, tandis qu'un kilogramme (kg) est l'unité de base de masse dans le système international d'unités (SI).
2. 1 kilogramme est approximativement égal à 2,2 livres.

Donc, un kilogramme de plumes est plus lourd qu'une livre de fer, car il correspond à environ 2,2 livres de plumes.