Mistral Large 4 Open Weights: What Le Chonk Means for You

On October 29, 2026, French AI lab Mistral did something most frontier labs won't do: it gave away the weights of its biggest model. Mistral Large 4, nicknamed "Le Chonk," is now downloadable from Hugging Face — a 1-trillion-parameter, natively multimodal model built for coding and cybersecurity. Anyone can inspect it, fine-tune it, and run it on their own servers, under a modified MIT-style license.

If that sounds like inside baseball, here's why it matters to you: this is the first time a Western lab has open-released a model at this scale with cyber defense as its headline job. It won't replace ChatGPT on your phone, and it won't run on your laptop. But for developers, security teams, and anyone who cares about who controls powerful AI, Mistral Large 4 is a genuine shift — and the details are more interesting than the headline.

Quick answer: what is Mistral Large 4?

Mistral Large 4 is Mistral AI's flagship open-weight model, released October 29, 2026, three weeks after its October 6 public preview. It has roughly 1 trillion total parameters (about 52 billion active per token, including embeddings), a 1-million-token context window, and native text-and-image input. It is pitched at coding agents and cyber defenders, scores 41% on Terminal-Bench 4.0 and 51% on the Artificial Analysis Cyber Index according to Mistral, and costs $0.70 per million input tokens via Mistral's API. The weights are free to download — but running them yourself needs data-center hardware.

What "open weights" actually means

"Open" gets thrown around loosely in AI, so let's be precise. With Mistral Large 4, Mistral published the model's weights — the trained numbers that make it work — on Hugging Face under the Mistral Ecosystem Open-Weight License, a modified MIT license. That means you can download, modify, fine-tune, and commercialize products built on it for free.

There is one important catch: if you want to offer Mistral Large 4 itself as a hosted model service (model-as-a-service), you need a separate commercial license from Mistral. Building an app on top of it is fine; reselling the model as a service is not covered by the free license. That's a narrower restriction than a pure open-source license, but far more freedom than a closed API-only model, where you can never see or move the model at all.

The practical payoff is control. A bank, a hospital, or a government agency can run Large 4 inside its own firewall, on its own data, without sending sensitive code or documents to someone else's servers. That sovereignty argument — European infrastructure, trained in Mistral's own data centers — is central to Mistral's pitch, and it's why partners like Booz Allen, Elastic, and Europe's cybersecurity agency ENISA publicly said they have been testing it.

Mistral Large 4 specs at a glance

Total parameters: ~1 trillion (mixture-of-experts: 128 routed experts, 4 active per token plus 1 shared expert)
Active parameters per token: ~52 billion including embeddings and output layers (~49 billion excluding them — both figures appear in Mistral's own materials, which explains the different numbers you may see quoted)
Context window: 1 million tokens — enough for very large codebases or long documents in one go
Input: natively multimodal — text and images, via a 1.6-billion-parameter Pixtral vision encoder
Reasoning modes: Low, Medium, and High, so developers can trade cost against performance (High is the default)
Checkpoints: FP8 and NVFP4 quantized versions on Hugging Face, plus community GGUF quantizations
Serving: vLLM or SGLang for self-hosting; also on Mistral's API and Studio, OpenRouter, Microsoft Foundry, and NVIDIA NIM containers

One honest note on the numbers: the 49-billion vs. 52-billion active-parameter figures are not a contradiction, just different accounting. Most coverage quotes one or the other. What matters is the idea: only a small slice of the trillion parameters fires for each token, which is why a model this large can answer at usable speed at all.

How good is it? The benchmarks, honestly presented

Every number below comes from Mistral or its partners — independent evaluations will take time, and you should read them that way. Still, the pattern is consistent: this is a top-tier open model for coding, and a strong but not chart-topping one for cyber.

Benchmark	Mistral Large 4 score	What it means
Terminal-Bench 4.0	41% (up from 28.3% at preview)	AI agents working in a real terminal; Mistral says this leads open-weight models, just ahead of GLM-5.3
Artificial Analysis Coding Agent Index	48%	Effectively tied with GLM-5.3, while using 47% fewer tokens per task — cheaper to run
FrontierCode 2.0 (Cognition, private)	19.7%	End-to-end code quality judged on whether a maintainer would merge the change; Mistral says it leads open models here
Artificial Analysis Cyber Index	51%	Finding, reproducing, validating, and patching vulnerabilities; third among open models, behind DeepSeek and Kimi variants
B3 Agent Security Benchmark	97.1% attack resistance	Up from 93.3% at preview; how well the agent resists being manipulated
Malicious cyber-request refusal	96.6%	Mistral's own measurement; the highest refusal rate it measured among open models

Beyond code and security, Mistral reports 73.3% on FinWorkBench for finance workflows and 84.5% on BrowseComp, up from 75% at the software fine-tuning checkpoint. On the broader Artificial Analysis Intelligence Index, Large 4 scores 26 — the highest among Western open-weight models, but fifth among open weights overall, with Chinese models such as DeepSeek V4.1 Flash and GLM-5.3 ahead on the aggregate. That's the fair summary: best-in-the-West open, not best-in-the-world open.

The three extra weeks of reinforcement learning between the October 6 preview and the October 29 weights release did real work — Terminal-Bench jumped 13 points. That also tells you something about how these releases work now: the "preview" is a public testing period, and the weights release is the improved, hardened version.

Why cybersecurity is the headline — and the controversy

Mistral is deliberately positioning Large 4 as a defender's tool: find vulnerabilities in your own code, reproduce them, patch them without breaking existing features, and audit systems at scale. ENISA says it is testing the model to identify vulnerabilities and audit code; Booz Allen says it matched other frontier cyber models on its internal source-code security benchmark with what it calls a compelling cost curve. Cisco and Cloudflare backed the open approach in Mistral's announcement.

But Mistral was unusually direct about the other side of the coin. During the preview, it says it monitored cyber-related use and found "a potential for uplift to conduct real attacks" — in plain language, the model could help a real attacker, not just a defender. Its response is twofold: the public model keeps guardrails that refuse an average of 96.6% of malicious cyber requests, and a specialized penetration-testing and red-teaming variant goes only to vetted enterprise partners, such as Orange Cyberdefense, under contract.

Here's the uncomfortable truth about open weights: once the weights are downloaded, they cannot be recalled, and a determined user can try to fine-tune the refusals away. Mistral knows this — Reuters quoted its VP of Science Pierre Stock on exactly that point. The company's bet is that the defensive value of putting a strong cyber model in every security team's hands outweighs the risk of misuse, and that keeping such models closed only hands the advantage to attackers who already have them. You can disagree with that bet, but it is at least an honest one, stated in the open.

How to use Mistral Large 4: three ways in

1. The API (easiest for most people)

If you just want to try it, start here. Mistral serves it as mistral-large-4 through its API and Mistral Studio, and it is also available through OpenRouter, Microsoft Foundry, and NVIDIA-hosted endpoints. Pricing is $0.70 per million input tokens, $0.07 per million cached input tokens, and $2.10 per million output tokens — roughly half the preview price, and competitive for a frontier-class coding model. If you build coding agents, pair it with careful prompt design: the model is strong, but agent results still depend heavily on how you structure tasks, which is where a tool like our AI Prompt Generator can help you draft and refine structured prompts before you spend tokens.

2. Mistral's own products

Large 4 is also available in Vibe, Mistral's coding assistant, where some users already had access during the preview. This is the no-setup route if you want the model inside a coding workflow rather than raw API calls.

3. Self-hosting (for organizations, not home PCs)

Download the weights from Hugging Face (listed as Mistral-Large-4-1T-A52B) and serve them with vLLM or SGLang. But mind the hardware catch: even the quantized version reportedly needs four NVIDIA B200 or B300 data-center GPUs. The full-precision model needs more. This will not run on a gaming PC, and that is by design — a trillion-parameter model is infrastructure, not an app. Reports say Mistral plans smaller distilled versions, including one that runs on a laptop, in the coming months, but no date has been confirmed. If your team is weighing self-hosting, that smaller version is the one to watch.

Step-by-step for the API route:

Create an account on Mistral's developer platform (Mistral Studio) and generate an API key.
Select the mistral-large-4 model in your request.
Start with the default High reasoning mode for hard coding or security tasks; drop to Medium or Low for simpler jobs to cut cost.
For code review, paste the code plus a precise instruction (what to look for: vulnerabilities, bugs, or both) rather than a vague "review this."
Always verify the model's patches yourself — run your tests before merging anything it suggests.

Mistral Large 4 vs. the open-weight field

Model	Size / approach	Where it leads	Access
Mistral Large 4	~1T total, ~52B active	Western open-weight leader; coding + cyber defense focus	Open weights + API ($0.70 / $2.10 per million tokens)
GLM-5.3 (Zhipu)	Open-weight rival	Effectively tied on coding index; slightly ahead on aggregate open rankings	Open weights + API
DeepSeek V4.1 Flash	Open-weight rival	Ahead on aggregate intelligence index and cyber index (57)	Open weights + API
Muse Spark Open (Meta)	816B total, 42B active	Developer coding on Apple Mac clusters; needs ~493GB even compressed	Open weights
Gemini 4 Carbon (Google)	Closed frontier model	Cyber benchmark leader, but limited to trusted defenders initially	Restricted rollout, not open weights

The comparison that matters most is the philosophical one. Google's Gemini 4 Carbon tops cyber benchmarks but is rolling out only to vetted defenders; Mistral's Large 4 is weaker on the cyber index but downloadable by anyone. One approach trusts gatekeeping, the other trusts distribution. The next year will show which one actually makes software safer — and security teams are now, for the first time, able to test both sides of that argument with frontier-class tools.

This also continues a pattern we have tracked all month on this blog: the open-weight frontier is moving fast. We covered Kimi K3.1 Pro, GPT-6.1 Terra, Mercury 3, and MiniMax M3.1 Flash in recent weeks — and Large 4 is the first of them to make cyber defense, rather than general chat or coding alone, its defining pitch. If you want the bigger picture, our homepage collects the latest AI model coverage, and our About page explains how we test and verify these claims.

Pros and cons

Pros

Truly downloadable frontier-class model — no vendor lock-in, full data control
Strong coding agent performance with 47% fewer tokens per task than its closest rival, so it is cheaper to run
1-million-token context window handles huge codebases
Adjustable reasoning modes let you control cost vs. quality
European, sovereign-AI option for organizations that cannot depend on US or Chinese providers

Cons

Cannot run on consumer hardware — self-hosting is data-center territory
Not the top open model overall; Chinese rivals lead on aggregate and cyber-index scores
All headline benchmarks are vendor-reported so far; independent validation is pending
Open weights cannot be recalled, and guardrails can potentially be fine-tuned away
Model-as-a-service use needs a separate commercial license — read the terms before building a hosted business on it

Who should actually care?

Security and platform teams: you can now trial a frontier cyber-capable model inside your own infrastructure. Start with the API, benchmark it on your own code, and only consider self-hosting if data residency demands it.

Developers building coding agents: the token efficiency is the real story. A model that ties the leader while using nearly half the tokens changes the unit economics of an agent product.

Everyone else: you probably won't use Large 4 directly, and that's fine. Its importance is structural: it proves a European lab can play at the open frontier, it pressures closed labs to justify their restrictions, and it gives every security team a new tool to test. The smaller distilled version, if it arrives as reported, is the one that could eventually reach your laptop.

Verdict

Mistral Large 4 is not the best open model in the world, and Mistral doesn't claim it is. It is the best Western open model, the most cyber-focused one, and — crucially — one you can actually download today. For coding agents it is a genuine contender; for cyber defense it is a serious new option with honestly disclosed risks. Try it through the API first, read the license before building a business on it, and don't buy four data-center GPUs just to say you self-host it. The open-weights race just got a lot more interesting, and Europe just announced it intends to stay in it.

Frequently asked questions

Is Mistral Large 4 free?

The weights are free to download and use under Mistral's modified MIT-style open-weight license, including commercial use. Using Mistral's hosted API costs $0.70 per million input tokens and $2.10 per million output tokens, and offering it as a hosted model service needs a separate commercial license.

Can I run Mistral Large 4 on my PC?

No. Even the quantized version reportedly needs four NVIDIA B200 or B300 data-center GPUs. A home PC or gaming rig cannot run a 1-trillion-parameter model. Use the API instead, or wait for the smaller distilled versions Mistral has signalled.

What is "Le Chonk"?

"Le Chonk" is Mistral's nickname for Large 4, a joke about its size: at roughly 1 trillion parameters, it is the company's largest model yet. The name stuck in the official announcement and across developer coverage.

How is Mistral Large 4 different from the October 6 preview?

The October 29 release is an improved version after three more weeks of reinforcement learning and customer testing. Terminal-Bench 4.0 rose from 28.3% to 41%, and attack resistance on the B3 benchmark rose from 93.3% to 97.1%.

Is Mistral Large 4 safe to use for cybersecurity?

It is built for defensive work — finding and patching vulnerabilities on systems you own or are authorized to test. Mistral says it refuses 96.6% of malicious cyber requests, but it also disclosed a potential uplift for real attacks, so only use it on your own or authorized systems, never to probe systems without permission.

Where can I download Mistral Large 4?

From Hugging Face, listed as Mistral-Large-4-1T-A52B, served with vLLM or SGLang. It is also available through Mistral's API and Studio, OpenRouter, Microsoft Foundry, and NVIDIA NIM containers.

Last updated: October 30, 2026. Methodology: this article was compiled from Mistral AI's October 29, 2026 announcement, Reuters' coverage of the release, and independent reporting from TestingCatalog, Unite.AI, and Skimwire, cross-checked against each other; where sources disagreed (notably 49B vs. 52B active parameters), both figures and the reason are stated. All benchmark scores are vendor-reported unless noted, and independent validation is still pending. This article is not sponsored by Mistral AI or any company mentioned. Questions or corrections? Reach us via our Contact page.

Search This Blog

The AI Tools