Mistral Large 4: 1 Trillion Open Weights Explained

That last part is the news. A preview through an API is not the same as open weights. With an API, you're renting someone else's model, on their servers, under their rules, with their logging and their rate limits. With downloadable weights, you can run the model yourself, fine-tune it, put it behind your own firewall, and inspect what you're actually deploying. That's why open-weight releases matter to enterprises, researchers, and governments in a way that another API announcement doesn't.

But "open" here comes with an asterisk, and it's a big one. We'll get to that, because it's the detail most coverage skips past — and it's the detail that decides whether Large 4 is actually useful for your business model.

First, what shipped.

What Mistral actually released

Large 4 is a mixture-of-experts model. Instead of activating the whole model for every token, it routes each token to a small set of specialist experts. The headline numbers, from Mistral and its Hugging Face model card, are:

• About 1 trillion total parameters, reported as 1.05 trillion in some assessments.
• 49 billion active parameters per token, excluding embeddings and output layers — 52 billion if you count those layers in.
• 128 routed experts, with four active per token plus one shared expert.
• A 1-million-token context window.
• Native image input alongside text, so it can take in diagrams, screenshots and documents visually, not just as extracted text.
• Support for deployment with vLLM or SGLang, the two serving frameworks most teams use for large open models.

Mistral says it trained Large 4 from scratch on 3,800 Nvidia Grace Blackwell GPUs in its own datacentres in Europe. Some secondary reports quote 4,000 GPUs, so treat the exact GPU count as approximate — the important point is that this was a large, European-based training run, not a fine-tune of someone else's checkpoint.

The released version isn't identical to the preview. Mistral says it applied three additional weeks of reinforcement learning, plus feedback from customers and security partners, before publishing the weights. That extra training is where most of the reported benchmark gains come from, which we'll look at in a moment.

You can find the checkpoint on Hugging Face as Mistral-Large-4-1T-A52B. It's also served through Mistral Studio, the Mistral API, OpenRouter, Nvidia-hosted endpoints as a NIM container, and Microsoft Foundry, which Mistral describes as the first cloud to carry it.

The licence catch, in plain English

Here's where you need to read carefully.

Large 4 is released under the Mistral Ecosystem Open-Weight License, which Mistral describes as a modified MIT licence. In practice, that means you can use, modify, fine-tune and commercialise products built on the model, for free.

Except for one carve-out: if you run it as a model-as-a-service business — that is, if your business is offering Large 4 itself as a hosted API to other people — you need a separate paid licence from Mistral first. You can own the weights. You cannot resell them as an API without permission.

That distinction matters more than it first sounds. A company building an internal coding assistant, a cyber-defence tool, or a document-processing product on top of Large 4 is in a very different position from a company whose whole pitch is "we host Large 4 cheaper than Mistral does." The first is clearly allowed. The second needs a commercial conversation first.

There's also a grey area that Mistral's announcement doesn't fully define: where does a SaaS product that wraps the model end, and a model-as-a-service offering begin? If your product is mostly Large 4 with a thin interface, get legal advice before you launch. If your product is a real application where the model is one component among many, you're on much firmer ground. Either way, don't assume "open weights" means "no restrictions at all." It doesn't, for this model.

What it costs to use, not just to download

Downloading weights is free. Running them is not.

A trillion-parameter model, even a sparse one that only activates 52 billion parameters per token, still needs serious hardware. You'll need multiple high-memory GPUs, fast interconnects, and someone who knows how to serve a mixture-of-experts model efficiently. Mistral ships a separate NVFP4 build for Blackwell serving, which helps with efficiency, and the model includes an embedded DSpark speculative-decoding head, which is designed to speed up generation. Even with those optimisations, self-hosting Large 4 is an infrastructure project, not a weekend experiment on a single gaming GPU.

For most teams, the API is the practical starting point. Mistral cut API pricing sharply for the weights release: $0.70 per million input tokens, $0.07 for cached input, and $2.10 per million output tokens. At preview in early October, those rates were $1.36 and $4.18 — so the release price is roughly half the preview price.

That pricing puts Large 4 in an interesting spot. It's not the cheapest open model you can call, but it's priced like a model Mistral wants people to actually use at scale, for coding and security work, rather than just benchmark and forget.

The benchmarks: impressive, with asterisks

Mistral's pitch is that Large 4 is the best open-weight model for coding. The numbers it points to, mostly from Artificial Analysis and Cognition's private evaluations, are worth looking at — with the usual vendor-benchmark caution.

Artificial Analysis gives Large 4 at its high reasoning setting 26 points on its Intelligence Index v5, up from 25 for the preview checkpoint. That makes it, in Artificial Analysis's assessment, the highest-scoring model with available weights developed outside China. It also reports that the updated model costs $2.06 per Intelligence Index task, down from $5.54 for the preview — a 63% reduction in cost per task.

On coding specifically, Mistral reports 48% on Artificial Analysis's Coding Agent Index, effectively tied with China's GLM-5.3, the top open model there, while using 47% fewer tokens per task. On Cognition's private FrontierCode 2.0 test, which asks whether a repository maintainer would actually merge the agent's change, Mistral scores 19.7%, against 19.3% for DeepSeek V4.1 Flash and 17.8% for GLM-5.3. On Terminal-Bench 4.0, Mistral reports 41%, up from 28% at preview.

Those are genuinely strong coding results, and the token-efficiency point matters: a model that gives you a similar answer for half the tokens is cheaper to run, even before you look at the headline price per token.

The cyber story is more complicated, and more interesting. Mistral is pitching Large 4 hard at cyber defenders. It says the model scores 51% on the Artificial Analysis Cyber Index, and that in its own tests it refused 96.6% of malicious cyber requests. At the same time, other coverage of the same release notes that on the cybersecurity index Mistral is showcasing, two Chinese open models score higher. So "best open model for coding" and "best open model for cyber" are not the same claim, and you shouldn't collapse them into one.

Mistral is also unusually candid about risk. It says it saw "a potential for uplift to conduct real attacks" during the preview, and that it is tightening guardrails in response. A more aggressive pen-testing version stays restricted to vetted partners, not in the public weights. That's the right call, but it also tells you something about the capability here: this is a model that security teams will want, and attackers will want to misuse, and the difference will often come down to who has better controls around it.

A full technical report is still, in Mistral's words, "coming soon." Until it lands, treat the benchmark figures as Mistral's and its partners' numbers, not as independent, peer-reviewed results.

Who should self-host, and who should just use the API

Self-hosting Large 4 makes sense if you have a clear reason to own the model: data that can't leave your network, a regulated environment, a need to fine-tune on proprietary data, or a workload large enough that API margins start to hurt. It also makes sense if you're a research lab, a government team, or a security vendor building a product where model control is part of the value proposition.

It does not make sense as a default. If you're a small team, a startup, or a company just trying to add a coding assistant to an existing product, start with the API or a managed endpoint through Microsoft Foundry, Nvidia, or OpenRouter. Measure quality, latency and cost on your real tasks first. You can always move to self-hosting later, and you'll move with actual data about whether Large 4 is the right model for you, rather than with enthusiasm for a trillion-parameter headline.

There's a middle path that's often overlooked: use the API to prototype, then fine-tune a smaller, task-specific variant for production. Mistral says customers can tune Large 4 into smaller variants for well-defined security jobs. That's a sensible pattern for cyber work, where a specialised model for triage, log analysis, or vulnerability explanation may be more useful — and much cheaper to run — than a general trillion-parameter model for everything.

What this means for the open-model race

Large 4 lands at a moment when the open-weight race is no longer just about who has the biggest model. It's about who has a model you can actually deploy, afford, and trust in a specific job.

On that measure, Mistral's release is significant for Europe in particular. It's a frontier-class open model trained in Europe, on European infrastructure, from a European lab, with weights you can download and run yourself. For organisations that care about sovereignty, data residency, or simply not depending entirely on US or Chinese model providers, that's a real option that didn't exist at this level a few months ago.

It also sharpens the competition. DeepSeek and GLM are still formidable on coding and cyber benchmarks, and in some cyber measures they score higher. OpenAI, Anthropic and Google still lead on closed, flagship capabilities. What Large 4 does is raise the floor for open models: if you can download a trillion-parameter model that codes this well, at this price, the argument for paying flagship prices for routine coding work gets harder to make.

That doesn't mean closed models are finished. For the hardest reasoning, the most sensitive safety work, and the most polished products, closed labs still have advantages. But for a growing slice of everyday developer and security work, the open option is now good enough that "good enough" is no longer a compromise — it's a choice.

The bottom line

Mistral Large 4, Le Chonk, is exactly what its nickname suggests: big, a bit unwieldy, and hard to ignore.

The weights are real, the coding results are strong, the API price has been cut roughly in half from preview, and the 1-million-token context window makes it practical for large codebases and long documents. The licence catch for model-as-a-service businesses is real too, and anyone planning to resell Large 4 as an API needs to deal with that first.

If you build software, run a security team, or care about owning the models you depend on, Large 4 is worth a serious trial — on your own code, your own logs, and your own tasks, not just on Mistral's benchmarks. Download it if you have the hardware and the reason to self-host. Use the API if you don't. Either way, test it against the model you're using today, and let that comparison, not the trillion-parameter headline, decide whether Le Chonk earns a place in your stack.