Mistral Large 4 is Mistral's new flagship open-weight model, now in public preview on the Mistral API, with the weights due by the end of October. For teams building AI agents, it is a strong candidate for workloads that must run on infrastructure they control, but the preview is the time to test it, not to move production onto it.
What was released
Mistral Large 4 was announced by Mistral on 6 October 2026 in the post Introducing Mistral Large 4, and listed the same day in the Mistral changelog as available in Public Preview under the model ID mistral-large-4. You can try it in Mistral Studio and call it through the API today.
The weights are not out yet. Mistral says it will release them by the end of the month, along with more detail on the architecture and its post-training method, and that until then it is red-teaming the model with security partners and state authorities.
What actually changed
Comparing the Mistral Large 4 model card with the Mistral Large 3 model card shows four practical differences:
- A bigger model. Mistral Large 3 has 41B active and 675B total parameters. Mistral Large 4 lists 52B active and 1.05T total parameters, plus a 1.6B vision encoder. The announcement gives a slightly lower active-parameter figure, so expect the final numbers with the weights.
- A longer context. The context window grows from 256k tokens on Large 3 to 1M tokens on Large 4.
- A higher list price. Per token, Mistral Large 4's list prices for input and output are well above Mistral Large 3's, and Large 4 adds a separate, much lower rate for cached input. For a short launch period, the changelog says Mistral is charging half the list price.
- One model for instruction and reasoning. Mistral describes Large 4 as a hybrid instruct-and-reasoning model with multimodal input, trained on data in more than 160 languages.
The API features are the same as on Large 3: structured outputs, function calling, document Q&A, batching, and the Agents and Conversations endpoints with built-in tools. That matters for migration, because an agent already using those features on Large 3 should only need a new model ID to start testing.
The announcement also reports Mistral's own scores on coding, agent, cybersecurity, finance, legal and vision benchmarks, including comparisons with closed models. These are vendor results on a preview that Mistral says is still being trained, so treat them as a reason to evaluate, not as a result for your workload.
What it means for teams building AI products and agents
In CodeDTX's view, the release matters most for teams that need a capable model they can eventually host themselves or keep in Europe.
- A real self-hosting option at the top end. Once the weights are out, Large 4 gives a model with long context, vision and tool use that you can run on your own hardware. That is a large serving footprint, so weigh it against the trade-offs in should you run an AI agent on a model you host yourself before buying GPUs.
- Data location without a new vendor. Mistral says the model will be offered in several regions, including a European deployment it runs end to end under European law. Teams with residency requirements should check which region their requests use; our guide to data residency for AI agents covers the questions to ask.
- Budget for the list price, not the launch price. The discount ends shortly after launch, and Large 4 costs more per token than Large 3. A longer context window also makes it easier to send more tokens per call. Use cached input for repeated system prompts and measure cost per task, as in controlling what an AI agent costs to run.
- Security teams get a model that will do security work. Mistral positions Large 4 for vulnerability research and incident response, including tasks where some closed models refuse. That is useful for defenders, but an agent that can reproduce exploits should only run against systems you own, in a sandbox, with a person approving changes; see how to stop an agent taking a wrong action.
When not to switch yet
Keep production agents on a generally available model while Large 4 is in preview: Mistral says the model is still improving, so its behaviour can change under the same ID. Wait for the published licence before planning to self-host, because the model card does not state one yet, while Large 3 is under Apache 2.0. And if Large 3 or your current model already passes your evals at an acceptable cost, the higher per-token price needs a measured gain to justify it. Test with the method in what AI agent evals catch, then move in stages.
Frequently asked questions
What is Mistral Large 4?
Mistral Large 4 is Mistral's new flagship model, announced on 6 October 2026. It is a multimodal mixture-of-experts model that combines instruction following and reasoning in one model, with a very long context window and support for function calling, structured outputs and agents. It is in public preview on the Mistral API, and Mistral says the open weights follow by the end of October.
How do I try Mistral Large 4?
You can use it in the Mistral Studio playground or call the Mistral API with the model ID mistral-large-4. It works with the same endpoints as earlier Mistral models, including chat completions, batching and the Agents and Conversations APIs. An agent already built on Mistral Large 3 can usually be tested by changing the model name and running your evaluation set.
Is Mistral Large 4 open source?
Mistral describes it as an open-weight model, but the weights have not been released yet. Mistral says they will be published by the end of October 2026, with more details on the architecture. The model card does not yet state a licence, so check the terms when the weights arrive before you plan to host it yourself or build a product on it.
Should we move our AI agent from Mistral Large 3 to Large 4?
Not in production while it is a preview. Run both models on the same real tasks, compare success rate, cost per task and latency, and check tool-calling behaviour on your own traces. Large 4 costs more per token than Large 3 at list price, so switch only where the gain is clear, and keep the older model available for rollback.
Can Mistral Large 4 run in Europe?
Mistral says the model will be available in several regions, including a European deployment that Mistral operates end to end under European law, and that the preview runs in its own European datacentres. Once the weights are released, you can also host it on your own infrastructure. Confirm which region serves your requests before you send regulated data.
