Decided on purpose,
not by default
We build features and products on Mistral AI models, with replies your code can read, data settings chosen and written down, and models run on the API or on your own servers, so you know what happens to everything you send.
Thanks, we’ll be in touch soon.
{{ heroStatus }}
You’re in good company
A few of the companies we’ve worked with.
When the free key went live
Many features on Mistral AI models start on a key made for trying things. Each request, what your product sends, goes out on whatever plan that key was made on, and nobody looks at the training setting, which decides whether Mistral AI may use what you send to train its models.
The model was picked in a quick test. Sometimes it’s an early-access model, one Mistral AI marks Labs or preview and offers for trying. And it’s called by a moving name, one Mistral AI points to newer versions as it releases them.
Then it goes live as it was built. The moving name points somewhere new. A model reaches its retirement date, and nobody had read the list of deprecated models. And legal asks what’s kept, and what may be used for training.
The test did its job: it proved the idea.
In a product, each of those choices is made on purpose: the plan and the settings, each model’s status and version, and where it runs.
What we build
Each task matched to the model your tests choose, with its settings, replies, versions and keys set on purpose, for one feature or a whole product.
A model sized to each task
Each job given the smallest Mistral AI model your tests accept, and a larger one only where the results ask for it.
Data settings, chosen on purpose
The plan, the training setting and each model’s status read before customer data goes in, and zero data retention requested where you need it.
Fields your code can read
A schema for every reply, so your product takes its values from named fields rather than from the wording.
Actions checked by your code
Tools the model may ask to use, each request checked by your own code before anything runs.
A test set from your own requests
Real requests with the answers you expect, run whenever a model, a setting or a prompt is about to change.
Room to run it yourself
Models released with open weights, their files to run yourself, run on your own servers where the license allows, on the same tests.
Fixed versions, planned moves
Each model called by the fixed version your tests passed, its retirement date written into your plan.
Access split by workspace
A workspace and its own keys for each part of the product, each with a spending limit of its own.
Live use reviewed with your team
Failures, refusals, the cost of each request and the replies people mark, read with your team.
- Assistants that take actions in your systems?AI Agents
- Answers for customers and staff, from your own documents?AI Chatbots & Assistants
- Your product inside the assistants people use?MCP & Agent Integration
- Models trained on your own data, with their pipelines and monitoring?Data Platforms & MLOps
Settings that decide what’s kept
On the Mistral AI API, what happens to the data a request carries isn’t one fixed rule. It follows from a few choices: the plan, the model’s status, the kind of request, the training setting, and zero data retention, which Mistral AI grants on request. Each can change what Mistral AI may keep or use.
So we write down the setup for each part of your product before customer data goes in. We check the training setting, read each model’s status before it goes in, and request zero data retention where you need it. The two settings are separate controls, so we set both.
Step 1
A first prototype
A single request and reply, sent to a model offered for production, on the free plan, with the training setting on.
Training: May be used to train Mistral AI models.
Kept: For a limited time, to check for misuse.
Step 2
The training setting, off
Turned off in the organization’s privacy settings, which are kept apart from the assistant app’s.
Training: Not used to train Mistral AI models.
Kept: For a limited time, to check for misuse.
Step 3
A paid plan
Neither line changes. A paid plan is where zero data retention can be granted, on request.
Training: Not used to train Mistral AI models.
Kept: For a limited time, to check for misuse.
Step 4
Zero data retention, granted
Mistral AI reviews each request for it and may approve or deny it. Once granted, it shows in the privacy settings.
Training: Not used to train Mistral AI models.
Kept: Not beyond what’s needed to return the reply.
Step 5
An early-access model
A model Mistral AI marks Labs or preview, swapped in to try it. Its terms treat it differently.
Training: May be used to train Mistral AI models, whatever the setting.
Kept: For a limited time, to check for misuse: zero data retention doesn’t apply to early-access models.
Step 6
A stored conversation, file or batch
A feature that keeps a conversation, a file or a batch has to store it to work.
Training: May be used to train Mistral AI models, whatever the setting.
Kept: Until it’s deleted or the account closes: zero data retention doesn’t cover stored data.
A few choices decide what happens to each request. We write yours down before customer data goes in.
Feedback sent to Mistral AI, with the request and reply it’s about, may be used for training whatever the setting.
Read the replacement before you move
Mistral AI names each model two ways: a fixed version, which stays the same once the model is offered for production, and a moving name, which Mistral AI points to newer versions as it releases them. So a product calls the fixed version its tests passed.
An early-access model can change under its own name, and may be retired before it’s ever offered for production.
When Mistral AI deprecates a model, its models overview gives a deprecation date, a retirement date and a replacement, and the model stays reachable until it retires. The replacement can itself be an early-access model, which the terms treat differently, so before a move we read its status and terms as well as its answers. Each move runs on the test set, then switches.
For a product already on a deprecated model, the next step is plain: its replacement read and tried on the test set, and the move planned before the retirement date.
Start with the keys and settings you have now
The review covers the plans and workspaces in use, and who holds each key; the training setting, and zero data retention where it applies; each model in use, with its status, its version and its retirement date; moving names against fixed versions; which requests are single and which are stored; the schemas and tools; the test set, or the cases that could become one; where data is processed; which models could run on your own servers, and under what license; and where cost can come down.
Then the first changes, in the order we’d make them, starting from a test set.
What you receive is a written review and a proposal for that first step.
Questions worth asking
Where is our data processed?
It depends on the endpoint. By default, requests go to the global endpoint of the Mistral AI API, which commits to no particular place. Regional endpoints process requests in the European Union or the United States, at a higher price and with fewer features: stored conversations, batches and files aren’t offered there, and account settings, keys and billing may still be handled elsewhere. Mistral AI lists its main service providers publicly. We choose the endpoint with your security team, and write it down.
Can we run the same model on our own servers?
For a model Mistral AI releases with open weights, yes, under its license, on servers you run. Some models are offered only on the API. The files you download are a version of their own, so we run your test set on them before planning on it, and read the license with your legal team.
Should we always use the newest version?
Not in production. A moving name changes when Mistral AI moves it, so we call the fixed version your tests passed, and try the newer one on the test set before switching.
Will it work in the languages our customers use?
Your test set decides. The docs Mistral AI publishes list the languages its models support, and say they may also do well in others. We include real cases in each of your languages, and check each one before launch and after every change.
What does it cost to work with you?
It depends on the features you need, where their models run and the first step you choose. Before any work starts, you get a written proposal with the scope, the price or rate, and how you pay.
Can we hire developers for our own team?
Yes, through a dedicated team: Webair developers who stay with your product over time and work beside your own people.
Mistral AI is a trademark of Mistral AI. Webair is independent of Mistral AI, and isn’t sponsored by, endorsed by or affiliated with it.
Let’s build it better
Tell us the feature you have in mind and the setup it runs on today, and we’ll suggest where to begin, in a short call.
Thanks, we’ll be in touch soon.
{{ ctaStatus }}