Mistral Large 4: Europe's trillion-parameter open model

Mistral AI has opened the public preview of Mistral Large 4, a model of roughly one trillion parameters that the French company calls "the most powerful open-weight AI system outside China". Since October 6 it has been available through the Mistral Studio API, and the company promises to release the weights before the end of October so that anyone can download it and run it on their own servers.

It is the first model to come out of the €3 billion round Mistral closed in September, and it arrives at a time when the most capable open models —Kimi K3, GLM-5.3, DeepSeek-V4 and Qwen— are almost all Chinese. With it, Europe is looking for its own alternative that does not depend on providers who could cut off access.

What Mistral Large 4 is

Size: about 1.05 trillion parameters in total according to the model card (the announcement rounds it to "one trillion"), in a mixture-of-experts (MoE) architecture. Around 50 billion are active per token: the announcement says 49 billion and the model card says 52 billion.
Multimodal: it accepts text and images (with a 1.6-billion-parameter vision encoder) and answers in text.
Context: the model card lists one million tokens, although Artificial Analysis measures 524,000 on the preview API.
Languages: trained on more than 160 languages, including every official EU language.
Training: from scratch, in about two months, on around 4,000 NVIDIA Grace Blackwell GPUs in Mistral's own European datacenters, drawing about 10 MW. The reinforcement learning phase is still running, so results may change before the weights are released.

Pricing and availability

During the preview, the API costs $0.68 per million input tokens, $0.07 per million cached tokens and $2.09 per million output tokens, half the list price ($1.36, $0.14 and $4.18). Mistral has not said when the discount ends. It will also offer vetted developers, cybersecurity firms and government agencies a version with fewer restrictions and more cyber-defence capabilities than the public API.

The weights will arrive "by the end of the month"; The Next Web and VentureBeat report October 27. The licence is still unknown: Mistral Large 3 shipped under Apache 2.0, but VentureBeat points to a custom Mistral licence.

Performance: what Mistral says and what others measure

Mistral's figures are preliminary and come from the company itself:

Agentic coding (DeepSWE v1.1): 62%, against 61% for GLM-5.3 and 57% for DeepSeek-V4-Pro.
Finance (FinWorkBench): 67%, tied with DeepSeek-V4-Pro.
Visual grounding (DIOR-RSVG): 73%, ahead of GPT-6 Astra's 68% and Kimi K3's 55%.
Legal agent (Harvey): 15%, the best of the group, although such low scores across all models show how far they still are from handling legal work on their own.

The first independent measurement is more cautious: on the Artificial Analysis Intelligence Index it scores 38, behind GLM-5.3 (45) and Kimi K3 (44) and ahead of DeepSeek-V4-Pro (36). It is also very verbose: it used about 200 million output tokens to complete the index, against 67 million for GPT-6.1 Sol, which erodes much of its per-token price advantage. On the plus side, it generates about 116 tokens per second.

To compare it with its main open rival, we covered Moonshot AI's Kimi K3 in July.

The European pitch: sovereignty and cyber defence

Mistral says its customers want two guarantees: that no foreign country can cut off their access, and a clear record of how the model was trained. Large 4 can be run on your own servers or used through an API region of the customer's choice, including an EU region covered by European law. Mistral works with more than 125 large companies, such as Airbus, ASML and HSBC, and some of them helped train the model through its Forge platform.

Cybersecurity is the other big argument. "The cyber defense capabilities will enable enterprises and governments to defend themselves," said Mistral's chief scientist, Guillaume Lample. With open weights, a bank or a government could use the model to protect its networks without asking a third party for permission. CEO Arthur Mensch put it this way: "The narrative that Europe cannot compete is something that is not true."

Impact for development teams

For anyone building with LLMs, Large 4 adds a European option with good value in agentic coding, vision and European languages, plus the option of self-hosting once the weights are out. Self-hosting is not trivial, though: in an MoE of this size every expert has to sit in memory, so you need several high-end GPUs. For most teams, the realistic route will remain the API or a provider that serves it.

Practical recommendations

Test the preview with your own tasks (not just benchmarks) and measure cost per completed task, not per token, given how verbose it is.
Budget with the list price, not the preview price, until Mistral sets an end date for the discount.
Wait for the weights licence before planning a self-hosted deployment or a commercial product built on it.
If data must stay in the EU, check the API region and the data processing agreement.
For self-hosting, size GPU memory for the full model, not just the active parameters.

What to watch next

The weights release, expected on October 27, and its licence.
The technical details, extra benchmarks and safety testing Mistral promises to publish before then.
How independent scores change once reinforcement learning finishes.
The family of specialised models Mistral will build on top of Large 4.

Conclusion: Mistral Large 4 is not yet the best open model in the world by independent measures, but it is the first European model at this scale with a promise of open weights. If Mistral keeps to the schedule and the licence is reasonable, Europe will finally have an open, sovereign alternative to the big Chinese models.