Andy Garcia · freelance dev
FRENES
Back to the blog
AI & Sovereignty

Mistral Large 4 Is an Open-Weight Model You Cannot Download Yet: What a European CTO Should Check Before Sending It a Client File

October 7, 2026 6 min read

On Tuesday, October 6, Mistral released a public preview of Mistral Large 4. Very officially, Mistral calls it "le Chonk". It is a natively multimodal mixture-of-experts model with about 1 trillion parameters, trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own European datacenters, and the preview is served on that same infrastructure. List price: $1.36 per million input tokens, $4.18 per million output tokens.

Simon Willison's verdict that evening was fair. It is "certainly not a Fable-class model", but Mistral is back to being "maybe about 6 months behind the frontier". On the Artificial Analysis Intelligence Index it scores 38. Mistral Large 3, last December, scored 9.

I build a LegalTech on Claude. Our users upload documents they would not show a competitor. A European model, run by a European company, with weights I could host myself, is close to ideal on paper. So I read this launch as a procurement checklist, not a leaderboard. Five questions, five answers as of today.

Are the weights open?

Not yet. The announcement says "Weights drop end of this month." Until then, Mistral is red-teaming the model with "cybersecurity leaders, vetted partners, and state authorities, who will access the same model with reduced moderation and expanded cyber capabilities." TechCrunch describes public access as a "public guardrail endpoint", with weights expected in about three weeks, after safety testing. Artificial Analysis currently files the model as proprietary.

Even the spec is still moving. The blog post says 49 billion active parameters. The model card in Mistral's docs says 52 billion active and 1.05 trillion total, plus a 1.6 billion parameter vision encoder.

Under which license?

This is the question I care about most, and nobody can answer it yet. On the model card, the Weights field and the License field both read "Coming soon". The catalogue badge says "Open", with no license attached.

Mistral's own catalogue shows why that matters. Mistral Large 2.1, from November 2024, is listed under the Mistral Research License. Mistral Large 3, from December 2025, under Apache 2.0. Mistral Medium 3.5 carries a "Modified MIT" label. Three models, three different answers to the question that matters: can this run commercially, on our own servers, with client data?

The contrast came the day before. On October 5, Reflection AI announced Beam, a 501-billion-parameter open-weight model with 23 billion active, text only. Reflection says it "advances the Western open-weight frontier" and measures it against GLM 5.2 and Qwen 3.8-Max; TechCrunch framed it as a rival to the Chinese open models. Le Monde framed Mistral's launch the same way, as a model meant to close the gap with the best Chinese competitors. Beam's weights are not out either; Reflection says later this month. But Reflection named the license in its announcement: Apache 2.0. Mistral did not.

"Self-host" also needs a reality check. The model card's GPU memory column says N/A. My own arithmetic: 1.05 trillion parameters at 16 bits is about 2.1 TB of weights, about 525 GB at 4 bits, before a single token of context. That is a cluster decision, not a server under a desk.

Where does it actually run?

Mistral's pitch is sovereignty. The model "will be available across multiple regions worldwide, including a European deployment that Mistral operates end-to-end, independently of other digital service providers and under European law."

The docs add the fine print. The default endpoint, api.mistral.ai, is global, and "Mistral does not commit to a specific inference location for requests sent to this endpoint." The EU endpoint, api.eu.mistral.ai, bills at 1.1 times the list price. It does not serve Agents, Batch or the Files API, function calling is the only supported tool, and account data such as API keys, billing and usage analytics may still be handled outside the region. Zero data retention exists on paid plans, on request, and Mistral reviews each request.

The 10 percent is now standard. Anthropic's own API accepts two inference_geo values, "global" and "us", the latter at 1.1 times; there is no EU value, and regional endpoints on Amazon Bedrock and Google Cloud carry a 10 percent premium. OpenAI charges a 10 percent uplift on its data residency endpoints. Mistral's difference is not the price of a region. It is who operates the hardware and under which law, which is a legal argument before it is a technical one.

Is it cheaper?

Per token, clearly. Claude Sonnet 5.5 and GPT-6.1 Sol both list at $2 input and $10 output. Claude Opus 5.5 is $4 and $20. Mistral's docs pricing page currently lists Large 4 at a "Sale price" of half its list price, $0.68 and $2.09.

Per task, no. Artificial Analysis reports that Large 4 generated about 200 million output tokens to run its Intelligence Index, against a median of 81 million, for $1.13 per task. On the same index, Sonnet 5.5 at medium effort scores 40.8 for $0.59 per task, with about 31 million output tokens. GPT-6.1 Sol at low effort scores 42.1 for $0.13. Those figures use list prices. The sale would roughly halve Mistral's number, which puts it level with Sonnet 5.5 at medium effort, a configuration that scores higher.

The price per token is a sticker. The cost per completed task is the bill. A verbose model can carry an output price nearly 60 percent lower and still cost more.

What did the charts leave out?

Mistral's launch page carries about twenty benchmark charts. I went through them one by one.

The coding charts compare Large 4 only with open models, and it does not lead them. Kimi K3 scores 68 to its 62 on DeepSWE 1.1. GLM-5.3 scores 40 to its 28 on Terminal-Bench 4.0. DeepSeek V4 Pro scores 66 to its 59 on SWE-Atlas-QnA. The text names the two models its combined coding score beats.

Closed models are picked with care. GPT-6 Astra, OpenAI's $10 and $50 flagship, appears on seven charts. Claude Opus 5.5 appears once, on the Artificial Analysis Cyber Index, drawn with its safety blocks, since Mistral's argument there is that closed models refuse security work. The previous Claude Opus 5 shows up twice, and beats Large 4 in a blind human evaluation on code, 4.22 to 3.74.

The chart closest to my job is Harvey's Legal Agent benchmark, run by vals.ai: Large 4 at 15.8, GPT-6 Astra at 5.4, no Claude model at all. Artificial Analysis's own Legal Index, a broader composite, flips the order: GPT-6 Astra between 52.9 and 58.2 depending on effort, Large 4 at 37.0, with Sonnet 5.5 at medium effort at 43.7.

What appears nowhere: Claude Sonnet 5.5 and GPT-6.1 Sol, the two models priced closest to it, and the Intelligence Index score of 38. That is normal launch behaviour. It is also information.

What I am doing on Monday

Nothing moves in production. Three things go on the list.

First, if you test Large 4 with anything sensitive, send it to api.eu.mistral.ai, call models.list there first to confirm the model is served in that region, and request zero data retention before the first real document, not after.

Second, run it on your own evaluation set at both reasoning levels the API exposes, "none" and "high", and record cost per completed task, tokens included. Artificial Analysis already suggests it will surprise you.

Third, treat the license as a gate. Self-hosting is the real reason a European CTO would pick this model over an American API, and that plan cannot start until the license text exists. I put a reminder on October 31.

If the weights land under a permissive license and the model holds up on our documents, this becomes the most interesting European option for teams like mine. Until then, it is a promising API with a European address.

Sources

A project like this one?

I design and deploy products like this. Let's talk.

Let's talk