Sovereign Model
also called Sovereign AI model, National model
A frontier model you can run on your own infrastructure, under your own jurisdiction, instead of renting it from a foreign provider.
A model whose weights are downloadable and whose training lineage sits under a named jurisdiction, so an organisation or government can host, fine-tune and govern it itself rather than depending on another provider's API and its revocation, pricing or policy changes. Mistral AI's Large 4, whose weights were released on October 29, 2026, is the current reference point: a 1-trillion-parameter model trained from scratch on European infrastructure and pitched explicitly as a foundation for sovereign AI.
Open weights alone do not make a model sovereign. The test is the whole package: downloadable weights, a licence that permits self-hosting and commercial use, training provenance you can point to, and hardware you can actually run. Large 4 passes the first three and strains the fourth, since the official serving recipes target multi-GPU servers, not a laptop.
Read the licence before calling anything sovereign. Large 4 ships under the Mistral Ecosystem Open-Weight License, a modified MIT licence that is free to use, modify, fine-tune and commercialise, except that a business operating a model-as-a-service offering needs a separate commercial licence from Mistral. A ministry running it for its own staff pays no licence fee; a company reselling access to it does.
The numbers to keep attached to Large 4 are Mistral's own: about 1 trillion total parameters with roughly 49 billion active per token (52 billion counting embeddings and output layers), a 1-million-token context window, and text and image input. On vendor-reported benchmarks it scores 48% on the Artificial Analysis Coding Agent Index, effectively tied with GLM-5.3 while using 47% fewer tokens per task, and 41% on Terminal-Bench 4.0, up 13 points from the October 6 preview after three more weeks of reinforcement learning. No independent evaluation of the released weights was available at release, and the full technical report was still coming.
The ministry test
A public agency is told a model is sovereign because its weights are public.
Ask four questions before accepting the label: can we download the weights without approval, does the licence let us run them commercially for our own users, can we say where the model was trained, and can we serve it on hardware we own or rent in our jurisdiction? Large 4 answers the first three yes for in-house use and the fourth only with multi-GPU servers; the model-as-a-service carve-out is the clause that changes the answer for a reseller.
Calling a model sovereign because its weights are downloadable. If you cannot run it, cannot fine-tune it under the licence, or cannot say where it was trained, you have a download, not sovereignty.
- blogMistral Large 4 (Le Chonk) is now open weight ↗
Mistral's October 29, 2026 weights release: a 1-trillion-parameter model trained from scratch on European infrastructure, pitched as a foundation for sovereign AI. Benchmark figures are Mistral's own, not independently verified.
- reportMistral Large 4 open weights: specs, license, cyber claims ↗
Release-day reporting with the serving details: about 49 billion active parameters per token, 1-million-token context, the modified-MIT licence with a separate commercial licence for model-as-a-service, and $0.70/$2.10 API pricing. Flags that the technical report was still coming.
Open Weights · Open-source model
A model whose trained parameters you can download and run yourself.
Mixture of Experts · MoE
A big model that only uses part of itself for each token.
Model Provider
The company or platform you call to run inference.
Vendor Lock-in
How hard it would be to leave your model provider.
Checkpoint · Weights snapshot
A saved copy of a model’s weights at one point in training or post-training.