The Three Tiers: Open, Premier, and Labs

A Three-Tier Distribution System

Every Mistral model belongs to one of three tiers, which determine its access conditions, license, and deployment terms. This classification isn’t just catalog trivia: it dictates what your legal team needs to approve, the nature of your invoice, and who is responsible for service availability at 3 a.m.

Open Tier — Open Weights

Open Tier models have publicly available weights: you can download, inspect, modify, and deploy them freely. Files are published on Hugging Face under the Apache 2.0 license, which authorizes commercial use, modification, and redistribution. In practice, this means three linked freedoms: host the model on your own infrastructure, fine-tune it on your data, and operate without a single network call to Mistral.

As of April 2026, this tier includes most flagship models. You’ll find Mistral Small 4, the recommended default hybrid, and Mistral Large 3, the multimodal frontier model. Magistral Small 1.2 covers reasoning, while Devstral 2 serves as the code agent. On the audio side, Voxtral TTS handles speech synthesis and Voxtral Realtime handles real-time transcription. The Ministral 3 range, in its 3B, 8B, and 14B variants, completes the lineup for edge devices.

The benefits of this tier reinforce each other. Sovereignty is total since data never leaves your perimeter, which mechanically simplifies GDPR compliance and requirements for regulated sectors—a compliance file moves much faster when no transfer occurs. Costs change nature: they become predictable, tied to hardware rather than consumed volume, enabling an annual budget where usage-based billing demands constant monitoring. Fine-tuning lets you adapt the model to your domain and vocabulary. Finally, longevity is assured: the model you downloaded remains usable even if Mistral changes its commercial policy.

Premier Tier — The Managed API

Premier Tier models are accessible only via the Mistral API; their weights are not published. You send requests, Mistral runs the infrastructure; there is no download or self-hosting. Billing is per token consumed, and the offer comes with service level agreements and support, which matters once an application hits production.

As of April 2026, this tier includes Mistral Medium 3.1 as the multimodal frontier, Codestral for code completion, OCR 3 for document understanding, Voxtral Mini Transcribe 2 for audio transcription, Codestral Embed for code embeddings, Magistral Medium 1.2 for frontier reasoning, and Mistral Moderation 2 for safety.

Premier’s central argument is zero infrastructure: no GPUs to provision, no model upgrades to plan. Scaling follows your request volume without intervention on your part, and you benefit immediately from the latest versions and optimizations. Simplicity of startup also carries weight: an API key and a few lines of code suffice, whereas a sovereign deployment requires a project of its own.

Labs Tier — Experimental

The Labs Tier gathers models in the exploration phase: accessible, but without stability or longevity guarantees. Access is generally provided via API with restricted quotas, and the model may be removed or modified without notice. This category serves to test, evaluate, and prototype; it is discouraged for production, where an unannounced API change costs you in incidents. The current Labs model is Leanstral (March 2026), a formal proof agent in Lean 4.

How to Choose Between the Tiers

Read this table line by line based on your dominant constraint. If it’s confidentiality, the first line decides it alone; if it’s time to deployment, it’s the third.

The Hybrid Strategy

In practice, many organizations don’t have to pick a side and combine the two main tiers. They reserve Open for processing sensitive data, fine-tuned models, and anything that must remain under direct control. They rely on Premier for specialized functions—document extraction, transcription—where internal development makes no sense, and for the operational peace of mind a managed service provides.

This type of architecture is common: a self-hosted Open model handles client files, while a Premier service takes care of a peripheral, non-sensitive task. The exercise consists of drawing the boundary in the right place, according to your regulatory constraints and budget.

Key Takeaways

• Open = weights available, Apache 2.0, self-hosting and fine-tuning possible
• Premier = managed API only, pay-per-use, infrastructure managed by Mistral
• Labs = experimental, no guarantees, exploratory use
• Most of Mistral’s flagship models are in the Open tier, which is distinct compared to competitors
• A hybrid Open + Premier strategy is often optimal in enterprise