Mistral Large 4: from open weights to sovereign deployment
Mistral has opened a preview of a one-trillion-parameter multimodal model with 49 billion active parameters, trained and served from its own European data centres. For an enterprise, however, sovereignty will only be achieved after licensing, sizing, controls, and reversibility have been verified.
1. A preview, not yet an asset deployable everywhere
On October 6, Mistral launched the public preview of Mistral Large 4 through its API and announced that the weights would be released by the end of the month. The model combines instruction following, reasoning, multimodality, coding, and agentic use cases. Mistral says it trained the model from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its own European data centres, where the preview is also served.
The announcement further promises multi-region availability, including a European deployment operated end-to-end by Mistral under European law, and the ability to run in private cloud or on premises. Yet the weights, architecture details, further benchmarks, and post-training methodology have not been released. An industrialisation decision must therefore separate what is available today from what has been announced.
2. What changes for a Belgian or French company
For an SME, mid-market company, large enterprise, or administration, the concrete change is greater choice between a European API and controlled deployment. Teams can evaluate the same foundation for document RAG, coding agents, engineering drawing analysis, or cybersecurity tasks, then balance time to value, data control, and infrastructure cost.
Open weights do not make operations lightweight. A model of this size requires careful assessment of memory, parallelism, latency, energy, monitoring, and operational skills. An organisation should compare the full model with a specialised or smaller variant on its own data and languages, rather than confusing maximum capability with the best total cost.
3. Governing high-risk uses
Mistral highlights performance in cybersecurity, finance, law, and agentic workflows. These domains specifically require controls outside the model: least-privilege tool access, isolated environments, RAG document provenance, human validation, and exportable logs. The ability to reduce refusals for vulnerability research may help defenders, but it also raises the need for containment and monitoring.
For connections to Odoo or another ERP, start read-only. Each agent needs its own identity, tools approved by process, spending limits, and approval before accounting entries, purchases, payroll, or customer records can be changed. Benchmark quality proves neither the correctness of a business action nor the compliance of a processing activity.
4. Underside analysis: a sovereign chain still needs evidence
European training, independently operated European hosting, and prospective weight availability are three separate levers. The first indicates where capability originated, the second addresses jurisdiction and service dependencies, and the third creates an option to change operators. None is sufficient alone: licensing, hardware, runtime, telemetry, support, updates, keys, backups, and exit must all be documented.
The right architecture may be local, cloud, or hybrid. A serious pilot compares at least the European API and a controlled target using the same scenarios, metrics, and simulated incidents. It also confirms that prompts, RAG indexes, adapters, evaluations, and logs remain exportable so the model can change without rebuilding the whole system.
5. Evaluation plan
Wait for the weights and their licence before reaching a conclusion on self-hosting. Meanwhile, build a representative evaluation set in French and the languages actually used, measure accuracy, citations, security, latency, and cost, then test tool permissions and rollback. Sensitive use cases also require an independent review of outputs and contractual terms.
Priority: treat “open weight”, “hosted in Europe”, and “operated sovereignly” as three separate properties, each supported by technical, legal, and operational evidence.
Assess a sovereign AI architecture