Mistral AI API Platforms: Models and Access

Two ways to use Mistral models

Mistral has two distinct model tracks. Some models have downloadable weights under a stated licence, so you can assess whether to run them on infrastructure you control. Others are offered as hosted services: you send requests to an API rather than deploying the model weights yourself.

For a team that wants to manage its own deployment, Mistral Small 4 and the Ministral 3 family are listed under Apache 2.0. Ministral 3 comes in 14B, 8B and 3B sizes and supports text and vision. The smaller sizes may be practical to evaluate where compute is limited, but test the actual workload before choosing based on size alone.

Hosted models cover tasks beyond chat. Mistral lists document OCR, transcription, speech generation, code completion, embeddings and moderation among its premier offerings. These can spare you from operating a specialised model, but you depend on the API route and terms available from the provider you choose.

Check the current catalogue before choosing

Model catalogues change, and older guides can leave retired names in circulation. Mistral’s official overview marks Magistral, Devstral, Leanstral and Mistral Large 3 as deprecated or retired. Don’t base a new integration on a third-party article that describes one of these as current; check the model card and the API’s available model names before you build around it.

The current overview lists Mistral Small 4 and Ministral 3 as open-weight options. It also lists Mistral Large 4 as a general-purpose multimodal model with open weights, but under the Mistral Ecosystem Open Weight License rather than Apache 2.0. For hosted specialist work, the overview lists Codestral (25.08) for code completion and OCR 4.1 and OCR 3 for document extraction.

Don’t treat everything displayed in a Mistral catalogue as made by Mistral. Its overview also includes partner models from other makers. If model origin matters to your procurement or product claims, check the individual entry rather than relying on the catalogue heading.

Read the licence model by model

“Open weight” describes access to model weights; it does not mean every Mistral model has the same licence or that all usage conditions are identical. Mistral Small 4, the Ministral 3 sizes and Shieldstral 1.0 are listed under Apache 2.0. Large 4 uses the separate Mistral Ecosystem Open Weight License. Review the applicable licence and model card before commercial deployment.

That check matters if you plan to modify a model, distribute it with software, or run it for customers. Confirm the exact version and its licence, then check whether your intended use meets the conditions. A hosted model is a different decision: you use the service under its provider’s terms, and you do not get weights to deploy yourself.

For privacy-minded teams, a European maker may be one reason to evaluate Mistral, but it is not proof of GDPR compliance, EU data residency or a no-training commitment. Read the relevant terms and ask the API platform how it handles request data. If your product processes sensitive information, include that review before sending real user content.

Choose by task, not just by chat quality

Mistral’s specialist models are useful when a product needs a defined capability rather than a general chat endpoint:

• Documents: OCR 4.1 and OCR 3 are hosted options for extracting text and layout from scanned files. The overview describes bounding boxes and confidence scores, which can help you locate extracted content and flag uncertain results for review.
• Meetings and calls: Voxtral Mini Transcribe 3 and Voxtral Small Transcribe 3 handle transcription and diarization. The realtime Mini Transcribe option is intended for live audio workflows, so measure end-to-end delay in your own call setup.
• Voice output: Voxtral TTS 3 offers speech generation, voice cloning and streaming. It is a hosted premier model, not an option to download and self-host.
• Safety: Mistral Moderation 2 is hosted and includes jailbreak detection; its listed context is 128K tokens. Shieldstral 1.0 is an Apache 2.0 open-weight alternative for text and image safety classification.
• Code completion: Codestral (25.08) is listed as a hosted code-completion model. Confirm that the route you intend to use actually exposes it.

Embeddings are also in the catalogue: Codestral Embed (25.05) and Mistral Embed (23.12). For search or retrieval, test them against your own documents and queries; a model name alone won’t tell you whether the results fit your data.

Direct API or a multi-model platform?

Going directly to Mistral gives you a route to its own hosted models and avoids an extra platform between your application and the model provider. It can be the simpler choice when you have already selected a Mistral endpoint and want to manage that integration yourself. Check current model availability, terms, billing and operational requirements before committing.

A multi-model platform can make it easier to test Mistral alongside models from other makers through a shared integration. Depending on the platform, you may also gain consolidated billing or routing and fallback controls. That can reduce the work involved in comparing alternatives, but it does not guarantee a lower total cost: platform fees, model availability and usage terms all matter. Compare equivalent requests and include any extra platform charge in your calculation.

Across all 7 plans from 7 providers in our full AI API comparison, monthly prices start at $35.00; these are full-comparison figures, not prices for the Mistral filter. We last synced prices on 31 October 2026. Use the comparison as a starting point, then check the provider’s current model list and pricing for the exact Mistral endpoint you need. A platform’s access to Mistral does not guarantee that it carries every model in Mistral’s catalogue.

What the filtered comparison can tell you

This page narrows our comparison to platforms listed as offering Mistral AI models. The live options shown here are AI/ML API, Eden AI and WaveSpeedAI. Their rows also list model families from other makers, which can be useful if you expect to compare or switch models without building a separate integration for each one.

Use the platform rows to compare the available service and then verify the specific Mistral model route. The page’s model-type filters include audio, speech, music and 3D, while its need filters include routing and fallback and open source. Those categories help you narrow the search, but they are not confirmation that a particular provider exposes Voxtral, OCR 4.1 or Codestral. Check the provider’s endpoint list, licence and terms before you plan an implementation.

If you need downloadable weights, compare the open-weight route separately from a hosted API listing. If you need speech, OCR or code completion without operating specialist infrastructure, ask whether the platform offers the precise hosted model and features you need. That small check can prevent a platform choice from becoming a rewrite later.