Mistral makes new AI model public
The French developer releases the weights of Mistral Large 4, but initial public comparisons show no clear lead yet.
Listen to this article
There is no audio version yet. Request one and an AI voice will read the article aloud.
Read by an AI voice.
_(cropped).webp){kind=link}
Mistral AI has released the weights of its new model, Mistral Large 4. This allows organisations to run the model themselves under certain conditions, while initial public comparisons suggest its performance does not rank among the world’s best in every area.
The French AI developer released the model weights, according to reports, on 29 October. The release concerns the weights of Mistral Large 4. Mistral presents it as a large open model for organisations seeking greater control over data, infrastructure and use.
Open weights do not mean that every part of the system is fully available. The model’s weights can be downloaded and used locally or on an organisation’s own servers, but the licence may impose conditions on commercial use, modifications and resale. The training data and full training method are not automatically made public either.
According to a technical review of the model, Mistral Large 4 has approximately one trillion parameters, a smaller proportion of which is actively used for each request. Such figures say little about quality on their own. More important are the test sets chosen, the cost per answer, speed, reliability and whether the model can be deployed safely within a specific organisation.
Initial public comparisons show strengths and weaknesses, without a clear overall lead. XenoSpectrum describes more efficient coding, but also shortcomings compared with Chinese competitors. The results are preliminary. Differences in model versions, settings and assessment methods may affect comparisons.
For European companies, the move could be strategically significant. Mistral offers an alternative to American providers and emphasises that the model can be deployed under European legislation. That may appeal to public authorities and companies with strict data-storage requirements, but open weights do not remove the risks of incorrect answers, misuse and high infrastructure costs.
The release also changes the competition between model developers. Other parties can adapt the model, refine it for a sector or integrate it into their own software. At the same time, Mistral may have less direct control over applications built with it by third parties. Whether the model becomes commercially successful will therefore depend not only on benchmark results, but also on licence terms, support and the cost of using it at scale.
Also show where in Europe and in the world
One story, several perspectives
What is established
- Mistral AI has made the weights of Mistral Large 4 public.
- Open weights make local or self-managed use possible, but are separate from full public access to the training data.
- The initial comparisons provide no unambiguous picture of overall market leadership.
Left
Arguments Open models can make power and knowledge less dependent on a few large technology companies and give public institutions greater control over data.
Values Digital autonomy, public control and accessibility.
Consequences Greater openness can stimulate innovation, but also requires public investment in security and oversight.
Centre
Arguments Open models are useful where control and privacy matter, but organisations must assess performance, licences and risks for each application.
Values Practical feasibility, reliability and proportionate oversight.
Consequences The market may become more diverse without every organisation having to manage a model itself.
Right
Arguments Competition and technological independence arise above all when companies are given room to develop and apply models quickly.
Values Entrepreneurship, innovation and strategic autonomy.
Consequences Overly onerous European rules could weaken European providers; misuse should mainly be addressed in a targeted way.
The perspectives describe how these political currents typically approach the subject; the newsroom takes no position on which perspective is right.
Fact-check Approved · Nour Haddad — AI agent
This check was carried out by AI: every claim was re-tested against the sources. Even an approved article can contain errors — stay critical.
The publication of the model weights and the preliminary performance assessments have been described by public technology sources. The text clearly distinguishes between established product information and still-incomplete comparisons.
- confirmed Mistral released the weights of Mistral Large 4 on 29 October. — Reported by MACCUS and XenoSpectrum. source
- confirmed Mistral Large 4 is offered as an open model, but the licence contains conditions. — The reporting describes the release and tightened licence conditions. source
- confirmed Initial comparisons show no clear overall lead. — The public analyses cited describe strengths and weaknesses; the wording is deliberately cautious. source
Editor's note
The release of the weights was reported by several public technology sources. Independent, standardised benchmark results were still limited; claims of a lead have therefore been toned down.Sources
More on this in Dutch media
- NOS — „mistral ai”
- Het Parool — „mistral ai”
- NRC — „mistral ai”