Mistral Large 4 Open Weights Are Live: How to Download and Run Le Chonk

On October 29, 2026, the wait ended. Mistral Large 4 open weights are now live on Hugging Face, which means anyone can download the French lab’s trillion-parameter flagship — nicknamed Le Chonk — and run it on their own hardware. Three weeks earlier, the model arrived as an API-only preview; today, the files themselves are public.

This is a follow-up to our original Mistral Large 4 guide, which covered the October 6 preview in full. That article explained what the model is, how it scored, and what it cost. This one covers only what is new: what you can actually download, the licence terms that govern it, the hardware it demands, the improved model’s fresh benchmark scores, and the new API pricing. If you want the full backstory — including why a trillion-parameter model is called a fat kitten — start with the original guide.

Table of Contents

What Just Changed: The Weights Are Out

Mistral Large 4 open weights running in a data center server room
Mistral Large 4 open weights need serious server hardware — a trillion parameters does not fit on a desktop.

Mistral promised the Mistral Large 4 open weights by the end of October, and it delivered with two days to spare. On October 29, the company published the weights of an improved version of Mistral Large 4 on Hugging Face under the repository name mistralai/Mistral-Large-4-1T-A52B. This is not simply the preview build with a download button attached: Mistral says the downloadable model went through three more weeks of reinforcement learning after the preview, with gains in coding, defensive cybersecurity, and reasoning efficiency.

The timing matters for US readers because the release closes the biggest open question from the preview. Until now, Large 4 was a model you could only rent through an API. Now it is a model you can own, audit, fine-tune, and run behind your own firewall — provided you have the hardware, which is a serious caveat we will get to below.

Reuters covered the release on the same day, framing it around cybersecurity: Mistral is pitching Large 4 directly at cyber defenders, and the downloadable weights are the mechanism that lets security teams deploy it without sending sensitive data to someone else’s cloud.

Mistral Large 4 Open Weights: What You Can Actually Download

Here is exactly what the Mistral Large 4 open weights download includes right now:

  • The full checkpoint in FP8 format. The main release ships as 54 files in 8-bit floating-point precision — roughly half the memory of the usual 16-bit format. The official serving recipe for vLLM, the production inference software most teams use, spreads the model across 8 GPUs.
  • A 4-bit NVFP4 quantized build. Alongside the full checkpoint sits a smaller quantized version that Mistral says runs on four Nvidia B200 or B300 GPUs with the same performance. Quantization shrinks the model by representing its numbers with fewer bits, trading a little precision for a much smaller footprint.
  • An embedded DSpark speculative-decoding draft head. This is a small helper model baked into the checkpoint that proposes blocks of three tokens at a time, which the big model then verifies. Mistral measured up to 1.9 times higher throughput from serving with the draft head than without it — a meaningful speed win for anyone running the model in production.
  • No approval step. Unlike some gated model releases, the download requires no application or waiting list. You can pull the files immediately.

There is also a softer on-ramp for the merely curious. Mistral opened a demo gallery that gives each visitor a small free budget during the first week, so you can try the model without downloading anything or paying for API calls. If you just want to see what Le Chonk can do, that is the cheapest way in.

One important distinction: the download is the public model. The more aggressive penetration-testing version with reduced moderation and expanded cyber capabilities — the one Mistral gave to vetted cybersecurity partners, partners, and state authorities during the preview — stays restricted. Do not expect the public weights to behave like that version, and do not build a security product assuming they will.

The Licence: Read This Before You Build a Business

Mistral Large 4 open weights cybersecurity and licence protection
The licence is the most important document in the Mistral Large 4 open weights release — read it before building.

The single most important document in the Mistral Large 4 open weights release is not a benchmark table — it is the licence. Mistral Large 4 ships under the Mistral Ecosystem Open-Weight License, a modified version of the MIT licence. Here is what it means in plain language:

  • You may use, modify, fine-tune, and distribute the model for free. Internal deployment inside your company is covered, and so are products with the model embedded for end users.
  • You may commercialize products built on it. Selling software that uses the model is permitted.
  • The one big carve-out: model-as-a-service. If your business — or an affiliate — operates a service that offers the model itself to customers (the classic hosted-API business), you need a separate commercial licence from Mistral before doing so commercially. This is the clause that separates “open-weight” from “open-source” in the purest sense, and it is the clause most likely to trip up a startup.

Why does this matter? Because the original preview article flagged the licence as the great unknown: Mistral’s earlier Large 3 shipped under Apache 2.0, and reporting suggested Large 4 would use a custom licence instead. Now we know. For most readers — researchers, internal enterprise teams, security departments, and product builders embedding the model — the terms are generous. For anyone planning to resell raw model access, the terms require a conversation with Mistral first.

Practical advice: read the licence file in the Hugging Face repository before you commit engineering time. Licence terms are the kind of detail that is boring until it is expensive.

Hardware: What It Takes to Run a Trillion Parameters

Mistral Large 4 open weights GPU hardware Nvidia Blackwell chips
Running Mistral Large 4 open weights takes four to eight Blackwell-class GPUs.

Let us be blunt: you cannot run the Mistral Large 4 open weights on a laptop, a gaming PC, or even a single high-end workstation GPU. A trillion parameters at 8-bit precision needs roughly a terabyte of GPU memory, which is why the official recipes call for server-grade hardware.

  • Full FP8 checkpoint: 8 GPUs, per the official vLLM recipe.
  • NVFP4 quantized checkpoint: 4 Nvidia B200 or B300 GPUs (Blackwell-class hardware), which Mistral says delivers the same performance.
  • Serving software: vLLM or SGLang, the two production inference engines the release targets.

For most US teams, the realistic paths are rented, not owned. You can rent Blackwell GPU clusters from a cloud provider by the hour, use a hosted endpoint, or skip self-hosting entirely and call the API. Microsoft Foundry is the first cloud provider carrying the model, with private deployment options for Azure developers, and Nvidia offers it as an optimized NIM container for teams already on Nvidia infrastructure.

The honest cost math: the weights are free, but the hardware is not. A four-to-eight-GPU Blackwell setup rented by the hour can still cost hundreds of dollars per day of heavy use. Compare that against the API price (below) before deciding that “free download” means “free to run.” For occasional use, the API wins. For sustained, high-volume, or data-sensitive workloads, self-hosting starts to make sense — and the privacy argument, keeping prompts and data inside your own network, is often the real reason teams choose it.

New Benchmarks: What the Improved Model Scores

Mistral Large 4 open weights coding benchmarks developer screen
The improved model behind the Mistral Large 4 open weights posts the top open-weight coding scores.

The downloadable model behind the Mistral Large 4 open weights is an improved build, and Mistral published fresh numbers for it. As with every launch figure, treat these as company-reported until independent labs reproduce them — but they are still the best public evidence of where the model stands.

  • FrontierCode 2.0: 19.7%. Mistral says this is the top score of any open-weight model on this hard coding benchmark.
  • Artificial Analysis Coding Agent Index: 48%. That puts it effectively tied with China’s GLM-5.3, the top open model on that index — while using 47% fewer tokens per task, which directly lowers serving costs.
  • Terminal-Bench 4.0: 41%. A big jump from the preview’s claimed 28.3% — and a reminder of how much three weeks of extra training can move a score. Independent runs of the preview measured lower (22.73%), so wait for independent runs of this build before treating 41% as settled.
  • Artificial Analysis Cyber Index: 51%. The headline cybersecurity number for the release.
  • Malicious-request refusal: 96.6%. In Mistral’s tests, the public model refused nearly all malicious cyber requests — the safety behavior you want in a downloadable dual-use model.
  • Cybench: 93%. Forty real security-competition challenges, nearly all solved — consistent with the preview’s already-strong showing.
  • Vulnerability reproduce-and-patch: 82%. Still the highest score of any model on this Artificial Analysis test, which asks the model to reproduce a real vulnerability and then fix it.

What changed versus the preview? The coding numbers moved up across the board, the cybersecurity story got sharper (a dedicated Cyber Index score and an explicit refusal rate), and the efficiency claim — same capability with far fewer tokens — is now central to the pitch. The context window remains 1 million tokens on Mistral’s model card, and the model still accepts text and images as input, with around 49 billion parameters active per token (52 billion including embeddings).

A full technical report is, in Mistral’s words, “coming soon.” Until it lands, the benchmark story rests on Mistral’s own reporting plus the independent scores that trickle in over the coming weeks.

New Pricing: API Costs Roughly Halved

Alongside the Mistral Large 4 open weights, Mistral cut the API price. The new rates are:

  • Input: $0.70 per million tokens (down from the preview’s $1.36 list price).
  • Cached input: $0.07 per million tokens — unchanged, and extremely cheap for repeated context.
  • Output: $2.10 per million tokens (down from $4.18).

That is roughly a 50% cut, and it matches the 50% launch discount that applied during the preview — except now it appears to be the standard rate rather than a promotion. For budgeting, plan around $0.70/$2.10 and treat any future discount as a bonus.

For context: the model is still far more expensive per token than a small efficiency model, but it is now priced well below the closed frontier models charging $4–$20 per million tokens, while being downloadable in a way those models are not. And once you self-host, your cost stops being tokens entirely and becomes hardware plus electricity — the whole point of open weights.

The Cybersecurity Angle: Why Defenders Get It First

Mistral is not marketing the Mistral Large 4 open weights as a general chatbot. It is marketing it as a defender’s tool, and the open-weights release is the delivery mechanism for that pitch. Security teams can now run a frontier-class model entirely on-premises: vulnerability reports, incident logs, and proprietary code never have to leave the building.

The practical security jobs it targets are concrete: reproducing a reported vulnerability from a description, writing and testing a patch, triaging security advisories, analyzing suspicious code, and assisting in controlled penetration tests. The 82% reproduce-and-patch score and 93% Cybench score are the evidence Mistral points to, and the 96.6% refusal rate on malicious requests is the safety counterweight.

There is a genuine tension here that honest coverage should name. Open weights cannot be recalled. Once the files are public, a determined user can fine-tune away safety behaviors — that is true of every open-weight release, and it is sharper here because the model is unusually capable at security work. Mistral’s answer is staged access: the public model refuses malicious requests at a high rate, while the unrestricted penetration-testing version stays with vetted partners. Whether that balance holds is something the security community will test in public over the coming months, and it is a story worth watching rather than a question this article can settle on day one.

How to Download and Run It: Step by Step

Mistral Large 4 open weights download setup for developers
Downloading Mistral Large 4 open weights is free — the hardware and bandwidth are the real costs.

If you have the hardware (or a cloud account), here is the practical path to the Mistral Large 4 open weights:

  1. Go to Hugging Face and open the official Mistral Large 4 repository, mistralai/Mistral-Large-4-1T-A52B. Confirm you are on Mistral’s official account — copycat repositories appear fast after big releases.
  2. Read the licence file first. Check the model-as-a-service clause against your planned use before you download tens of gigabytes.
  3. Choose your build. Take the FP8 checkpoint if you have eight GPUs available; take the NVFP4 quantized build if you have four Blackwell GPUs and want the smaller footprint.
  4. Serve it with vLLM or SGLang. Both are supported, and vLLM has the official multi-GPU recipe. The embedded DSpark draft head is there to boost throughput — enable it.
  5. Customize with Mistral Forge if you need fine-tuning or task-specific adaptation rather than raw inference.
  6. No hardware? Use the API or a hosted endpoint. Mistral’s API at $0.70/$2.10 per million tokens, Microsoft Foundry, or Nvidia’s NIM container all give you the same model without the GPU bill. The first-week demo gallery is free for a quick taste.

A note on download size: a trillion-parameter model in FP8 is on the order of a terabyte of files (54 files in the main release). Plan storage, bandwidth, and time accordingly — this is not a five-minute download, and a slow connection will make it a weekend project.

Mistral Large 4 Open Weights vs the Alternatives

How do the Mistral Large 4 open weights stack up against the other open-weight options US teams are actually choosing between?

  • vs. Chinese open models (DeepSeek V4 Pro, Qwen3.8 Max, GLM-5.3): The Chinese labs still lead the open-weight pack on broad aggregate benchmarks. Large 4’s claim is narrower and stronger: best open-weight coding scores, a tied Coding Agent Index with GLM-5.3 at far lower token cost, and the strongest cybersecurity profile of any open model. If your workload is coding or security, Large 4 is now the most credible non-Chinese option.
  • vs. Meta’s Muse Spark Open: Meta’s 816-billion-parameter open release is the other big Western open-weights story of the month, and our Meta Muse Spark Open explainer covers it in detail. Spark Open pitches efficiency and local-agent work; Large 4 pitches raw capability and agentic coding. Both are downloadable — the choice depends on whether you need maximum capability or maximum efficiency.
  • vs. the closed frontier (GPT-6, Claude, Gemini): The closed models still lead on general intelligence. Large 4’s pitch is not that it beats them head-to-head; it is that it gets close enough on coding and security while being downloadable, auditable, and runnable on your own hardware. For regulated industries and security teams, that trade is often worth more than a few benchmark points.
  • vs. just using the API: If you do not have Blackwell GPUs or a cloud budget, the API at $0.70/$2.10 is the sensible default — and for many teams it will stay the sensible default permanently. Open weights are an option, not an obligation.

Who Should Download It — and Who Shouldn’t

Download it if: you run a security team that needs on-premises AI; you operate GPU infrastructure and want a frontier-class coding model without per-token bills; you research AI safety, alignment, or model behavior and need full weight access; or you build products that embed a model and want to fine-tune it for your domain.

Skip the download if: you just want to chat with a smart model (use the API or demo gallery); you have a single consumer GPU (the model will not fit, full stop); or your business plan is reselling raw model access (talk to Mistral about a commercial licence first). For most individual developers, the API is the right door; the Mistral Large 4 open weights are for teams with infrastructure.

The Honest Caveats

  • All headline benchmarks are company-reported. Independent labs have not yet published runs of the improved build. The preview’s independent scores came in below Mistral’s claims on some tests, so keep that pattern in mind.
  • The full technical report is still pending. Architecture details beyond the headline specs will only be verifiable when Mistral publishes it.
  • The licence has a real commercial boundary. Model-as-a-service businesses need a separate licence. Read it before building.
  • Hardware is the price of admission. Free weights plus four to eight Blackwell GPUs is still an expensive setup, rented or owned.
  • Dual-use risk is real. A model this good at security work in public hands will be tested by attackers as well as defenders. Mistral’s refusal rate and staged access are mitigations, not guarantees.

Mistral Large 4 Open Weights FAQ

Are Mistral Large 4 open weights really free to download?

Yes. The weights are published on Hugging Face with no approval step and no download fee. Your costs are hardware, storage, bandwidth, and electricity — or API fees if you use the hosted route instead.

Is Mistral Large 4 open source?

Not in the strictest sense. It ships under the Mistral Ecosystem Open-Weight License, a modified MIT licence that permits use, modification, and commercial products — but requires a separate commercial licence for model-as-a-service businesses. “Open-weight with one commercial carve-out” is the accurate description.

Can I run Mistral Large 4 open weights on one GPU?

No. The FP8 checkpoint needs 8 GPUs per the official recipe, and the NVFP4 quantized build needs four Blackwell GPUs. Single-GPU users should use the API, a hosted endpoint, or the demo gallery.

What changed between the preview and the downloadable model?

Three more weeks of reinforcement learning, higher coding and cybersecurity scores, roughly halved API pricing ($0.70/$2.10 per million tokens), an embedded speculative-decoding draft head, and — obviously — the weights themselves.

Where do I download Mistral Large 4 open weights?

From Mistral’s official Hugging Face repository, mistralai/Mistral-Large-4-1T-A52B. Verify the account name before downloading, and read the licence file first.

Should I still read the original Mistral Large 4 guide?

Yes, if you want the full picture. Our original Mistral Large 4 guide covers the model’s architecture, the “Le Chonk” story, the preview benchmarks, and the broader open-weight race in depth. This article is the release-day update: what landed, what it costs, and how to run it.

Bottom line: the promise from October 6 is now real. Mistral Large 4 open weights turn Europe’s biggest open model from a rented API into a downloadable asset — with a licence worth reading, hardware demands worth respecting, and cybersecurity strengths that make it genuinely interesting for defenders. For teams with the GPUs, it is the most capable non-Chinese open model you can run today. For everyone else, the halved API price and the free demo gallery are the practical ways in.

Leave a Comment

Your email address will not be published. Required fields are marked *