This guide covers the 29 October 2026 open-weight release. Benchmark figures below are Mistral’s published results, not a test of your own software. A Model-as-a-Service business needs a separate commercial license.
1. What Mistral released on 29 October
Mistral had already offered Large 4 in public preview earlier in October. The 29 October release publishes the improved weights. Mistral describes a mixture-of-experts model with about 1.05 trillion total parameters, about 52 billion active parameters, a vision encoder, and a one-million-token context window. It accepts text and images and returns text.
The weights are on Hugging Face as FP8 and NVFP4 checkpoints. Mistral says you can try the model in Mistral Studio, call the Mistral API, use OpenRouter, or serve the weights yourself with vLLM or SGLang. NVIDIA-hosted endpoints and an NIM container are also listed. Microsoft Foundry is described as a managed route, including private deployment options for Azure.
Mistral also added three reasoning modes. Low is for simple tasks. Medium is for everyday tasks. High is the default and is meant for work that needs more care. Those labels are product settings, not a promise about accuracy.
2. API, a managed cloud, or your own servers
On small screens, scroll the table sideways to see every column.
| Route | What it is for |
|---|---|
| Mistral API | Mistral lists a price of $0.70 per million input tokens, $0.07 for cached input, and $2.10 per million output tokens. You send data to Mistral’s service. |
| Managed cloud | Foundry, OpenRouter, or NVIDIA hosting can be simpler than running the weights. Confirm the vendor, region, and retention terms before sending business records. |
| Your own servers | Download the weights and serve them with vLLM or SGLang. You keep the deployment, and you also keep the GPU, monitoring, and update work. |
Self-hosting is the point of an open-weight release when a bank, clinic, or public body cannot send customer records to a third-party API. It is a poor default for a small team that only needs a writing assistant. Large 4 is a very large model. Budget for a real serving stack, or stay on the API until a smaller build exists.
Mistral reports coding gains since the preview: 41% on Terminal-Bench 4.0, up 13 points, with about 30% fewer output tokens, and 48% on the Artificial Analysis Coding Agent Index while using about 47% fewer tokens than a tied open-weight peer. It also reports 51% on that group’s Cyber Index. Treat those as vendor benchmarks. Repeat one task from your own backlog before you adopt a score.
3. The license condition that changes a product plan
Mistral offers Large 4 under the Mistral Ecosystem Open-Weight license, which it describes as a modified MIT license. The company says you may use, modify, fine-tune, and commercialize the model, with one condition: a business that operates a model-as-a-service offering must obtain a separate commercial license before using it commercially.
Running the model inside your own company, or inside software you sell as an application, is a different case from reselling hosted model access. If your product is “customers call our API and we pass the prompt to Large 4,” read the license with counsel before launch. If your product is an Odoo workflow that calls a model you host for that one customer, write down who operates the service and who owns the records.
4. What to test before a rollout
- One checked task: pick a job you can grade, such as reading a non-sensitive vendor bill into fields a person reviews. Keep live payments, credentials, and customer files out of the first trial.
- Reasoning mode: start at Medium for everyday work. Move to High only when Medium fails a check you care about.
- Data path: record whether the prompt leaves your network. An API trial and a self-hosted trial answer different privacy questions.
- Defensive security only: Mistral describes stronger vulnerability triage and a high refusal rate on malicious cyber requests. Test security help only on systems you own or are contracted to assess. This article does not describe attack steps.
A useful pilot for an ERP team is a read-only summary of a module or a draft coding change that a developer reviews. The weights do not connect themselves to Odoo, your ledger, or your repository.
5. Frequently asked questions
Is Mistral Large 4 free to download?
Mistral published the weights for download and lists API prices of $0.70 per million input tokens and $2.10 per million output tokens. Download access is not the same as a free production deployment. Serving the model still requires infrastructure, and a model-as-a-service business needs a separate commercial license.
Can a small team run Large 4 on a laptop?
No. Mistral describes a model with about 1.05 trillion total parameters and tells developers to serve it with tools such as vLLM or SGLang. Use the API or a managed host unless you already operate that kind of infrastructure.
Do the published benchmark scores apply to Odoo work?
No. The coding and cyber scores are Mistral’s reported results on public benchmarks. Check the model on a task from your own system, with an answer you can verify, before you rely on it.