Mistral Large 4 is Mistral AI's new flagship model. Mistral announced a public preview on October 6, 2026, and it is available today through the preview API on Mistral Studio. The open weights are not out yet: Mistral says it will release them "by the end of the month," meaning the end of October 2026. The company also nicknamed the model "le Chonk," a joke about its size.
This guide covers what Mistral has actually confirmed: the model's size, how you can get access, the preview price, the benchmark results Mistral reports, and what is still unknown. Every number here comes from Mistral's launch post or its model documentation. Unless we say otherwise, benchmark figures are Mistral's own claims and have not been independently checked. The page reflects the situation on October 7, 2026.
What Mistral Large 4 is
According to Mistral, Large 4 is a mixture-of-experts model with 1 trillion total parameters and 49 billion active parameters per token. It is natively multimodal, meaning it takes images as well as text as input. Mistral's model card describes it as a hybrid "instruct-and-reasoning" model, so a single model handles ordinary chat, step-by-step reasoning and agentic tool use. Mistral calls it "our largest and most capable model to date."
A few other details Mistral has confirmed:
- Training hardware: trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own European datacenters. The preview runs on the same infrastructure.
- Languages: a large share of the training data was multilingual, covering more than 160 languages, including every official EU language.
- Version tag: the documentation lists it as version 26.10. Its predecessor, Mistral Large 3, is listed as 25.12.
- Still improving: Mistral says the reinforcement-learning run behind the preview is "still in flight," so the model you test today may change before the weights ship.
Mistral has not yet published the full architecture, context-window length or post-training details. It says those will come with the weights.
Release date: preview now, weights later
There are two release dates to keep apart:
- Preview API: available now (October 6, 2026). Anyone with a Mistral account can call the preview through Mistral Studio.
- Open weights: promised for the end of October 2026. Mistral has not given an exact day. Until the weights ship, Mistral says it is red-teaming the model with cybersecurity leaders, vetted partners and state authorities. Those testers get a version with "reduced moderation and expanded cyber capabilities."
That staged approach is similar to what other labs are doing with their most capable models. Google's Gemini 4 Argon also went first to vetted cyber defenders. The difference is the destination: Mistral says Large 4 will end up as downloadable weights you can run yourself, while Argon and GPT-6 Astra stay behind an API.
What license will it use?
Mistral calls Large 4 "open-weight," but the license has not been published yet. Mistral's model table lists it as "Open." For comparison, it lists Mistral Large 3 and Small 4 as Apache 2.0, and Medium 3.5 as a modified MIT license. Don't assume Large 4 will be Apache 2.0 until Mistral publishes the license text with the weights. If the difference between open-weight and open-source matters to you, read the license before you deploy.
Mistral Large 4 pricing
Mistral's launch page lists these preview API prices:
| Usage | Price |
|---|---|
| Input | $1.36 per million tokens |
| Output | $4.18 per million tokens |
Read aloud: one dollar thirty-six per million input tokens, and four dollars eighteen per million output tokens. These are preview prices and could change at general availability, so check Mistral's pricing page before you budget. Once the weights are out, you will also be able to self-host. In that case your cost is your own hardware or cloud bill, not a per-token fee. A 1-trillion-parameter model needs multi-GPU servers, not a laptop.
If you just want to chat with Mistral's models instead of using the API, Mistral's consumer assistant is Le Chat. Mistral has not said when Large 4 will be the default model there.
Benchmarks Mistral reports
Mistral's launch post has a lot of numbers. Here are the main ones, all reported by Mistral:
| Area | Benchmark | Mistral Large 4 |
|---|---|---|
| Coding | DeepSWE v1.1 | 61.7% |
| Coding | SWE-Atlas-QnA | 59.4% |
| Coding | Terminal-Bench 4 | 28.3% |
| Coding | Coding Agent Index | 49.8% |
| Agents | AutomationBench (657 business workflows) | 59.9% |
| Knowledge work | AA-Briefcase | 1,393 Elo |
| Cybersecurity | Cybench (40 competition challenges) | 93% |
| Vision | Dense 200 grounding | 42% (vs 41% for GPT-6 Astra) |
| Safety | Lakera B3 attack resistance | 93.3% of attacks resisted |
What Mistral says these numbers show:
- Coding: its Coding Agent Index score puts it ahead of DeepSeek V4 Pro and Qwen3.8 Max, Mistral says. In a blind human evaluation run with Surge AI, Large 4 ranked second of five models (3.74 out of 5), ahead of Kimi K3 and GLM-5.3 but behind Claude Opus 5 (4.22).
- Cybersecurity: Mistral says Large 4 ranks in the top five models globally on the Artificial Analysis Cyber Index. It also scores 82% on a test that asks a model to reproduce and then patch a real vulnerability, which Mistral calls the highest of any model.
- Knowledge work: in evaluations run by the third-party firm vals.ai on legal and financial tasks, Mistral says Large 4 beat GPT-6 Astra.
- Vision: it narrowly beats GPT-6 Astra on the Dense 200 visual-grounding test.
Notice what Mistral is not claiming. It does not say Large 4 beats the best closed models overall. Its main claims are that it is competitive with the strongest open models anywhere, and "significantly" ahead of any open-weight model built in the US or Europe. Where it compares directly against closed frontier models, the comparisons are narrow, such as specific vision, legal and finance tests.
The cybersecurity refusal debate
The most talked-about part of the launch is about refusals. Mistral says several leading closed models, including Claude Opus 5.5 and GPT-6 Astra, "score near zero" on the vulnerability-reproduction test because they refuse to do the task. Mistral argues that defenders often have to prove a flaw is real before they can fix it, and that refusals from closed models get in the way of that work.
The same post also claims Large 4 refuses malicious cyber prompts more often than any other open model. That is measured on refusal sets drawn from JailbreakBench, StrongREJECT and AgentHarm. Both things can be true: a model can refuse obviously malicious requests while still helping with legitimate vulnerability research. But open weights change the safety picture. Once weights are public, anyone can fine-tune away the safeguards. This is the central tension in open-weight releases of highly capable models, and the frontier labs' safety frameworks handle it differently. Our guide to the Preparedness Framework vs the Responsible Scaling Policy explains how OpenAI and Anthropic decide what is too risky to release.
On prompt-injection robustness, Mistral cites Lakera's public B3 AI Security Benchmark: Large 4 resisted 93.3% of attacks, which Mistral says is the highest score among its competitors.
Mistral Large 4 vs GPT-6, Claude and Gemini
"Mistral Large 4 vs GPT 6" and "vs Opus 5.5" are among the most common searches about the launch. The honest answer today is that no independent, apples-to-apples comparison exists yet. What we can say:
- Where Mistral claims a lead: a few narrow tasks. These are visual grounding (Dense 200), legal and finance tasks (vals.ai), and vulnerability reproduction, where the closed models mostly refused.
- Where it trails: in Mistral's own coding evaluation, it ranked below Claude Opus 5. Mistral makes no claim that it beats GPT-6 Astra or Claude across the board.
- What's different by design: self-hosting and data control. If your organization needs to run a frontier-class model on its own servers, or under European law with no US provider involved, Large 4 is aimed at you. GPT-6 Astra, Claude and Gemini 4 are API-only.
For the closed flagships, see GPT-6 Astra vs Claude Fable 5.1, Claude Fable vs Opus vs Mythos and our overall ChatGPT vs Claude vs Gemini vs Grok comparison.
Who should try it now
- Developers comparing open models: test the preview API now on your own prompts. When the weights land, you'll know whether self-hosting is worth it.
- Security teams: this is the clearest target audience. Remember that the version you can access is the standard preview, not the reduced-moderation build given to vetted partners.
- European organizations with data-sovereignty requirements: Mistral says it will run a European deployment end to end, under European law, independent of other digital service providers.
- Everyday chat users: there's no need to switch yet. Wait until Large 4 reaches Le Chat and independent reviews appear.
What to watch next
- The weights and license text, promised for the end of October 2026.
- Architecture, context-window and post-training details, promised with the weights.
- Independent benchmark runs once anyone can download the model.
- Specialized models built on Large 4, which Mistral says are coming.
We'll update this guide when the weights ship.
FAQ
When is the Mistral Large 4 release date?
The preview API launched on October 6, 2026. Mistral says the open weights will be released by the end of October 2026, but it has not given an exact date.
How much does Mistral Large 4 cost?
During the preview, Mistral lists $1.36 per million input tokens and $4.18 per million output tokens. Once the weights ship, you can self-host instead and pay only for your own compute.
How big is Mistral Large 4?
Mistral says it is a mixture-of-experts model with 1 trillion total parameters and 49 billion active parameters, trained on 3,800 NVIDIA Grace Blackwell GPUs.
Is Mistral Large 4 open source?
Mistral calls it open-weight, and the weights are due by the end of October. The license hasn't been published yet, so check the license text before assuming it is Apache 2.0 like Mistral Large 3.
Is Mistral Large 4 better than GPT-6?
Not overall, based on what Mistral itself claims. Mistral reports narrow wins over GPT-6 Astra on visual grounding and on some legal and finance tasks, and it ranked below Claude Opus 5 in Mistral's own coding evaluation. No independent head-to-head comparison exists yet.
