Mistral Large 4 Open Weights: What Small Teams Can Do

The short answer: you can download a trillion-parameter model now — but running it yourself is a data-center job

On October 29, 2026, Mistral AI published the open weights of Mistral Large 4, the model it nicknames “Le Chonk.” This is the same flagship it previewed through its API on October 6, now with three extra weeks of reinforcement learning behind it — and, crucially, with the weights themselves downloadable from Hugging Face instead of only rentable through Mistral’s API.

Here’s the honest summary for a small team: the download is real, the licence is genuinely permissive for most uses, and the API price got cut roughly in half. But self-hosting this model is not a weekend project on a spare workstation. The full-precision weights run to more than a terabyte, and the serving setups Mistral and Hugging Face document start at multi-GPU server nodes. For almost every small business, the practical choice is not “download versus buy” — it’s “rent Mistral’s API, rent the same weights from another host, or fine-tune and host through a provider.” This article walks through all three, with the numbers, so you can pick the one that fits your budget and your privacy needs.

One caveat up front, because it matters for every claim that follows: the benchmark scores in this article are Mistral’s own, reported in its announcement and echoed by launch coverage. They have not been independently reproduced by outside labs yet, and the model’s technical report is still to come. Treat them as a vendor’s best case, not a settled measurement.

What Mistral actually released

Mistral Large 4 is a natively multimodal mixture-of-experts model: about 1 trillion total parameters, with roughly 49 billion active for each token (52 billion if you count embeddings and output layers), and a 1-million-token context window. It takes text and images as input and produces text and code. A 1.6-billion-parameter Pixtral vision encoder handles the image side.

The release matters for three reasons that go beyond the parameter count — and independent launch coverage has usefully separated what Mistral announced from what outside analysts could verify so far. Mistral Large 4 open weights specs and licence Mistral’s own announcement is here: Mistral Large 4 (Le Chonk) is now open weight.

First, it’s the improved checkpoint, not just the preview. Mistral says the three weeks between the October 6 preview and the October 29 weights release were spent on additional reinforcement learning and customer testing. The headline improvement it cites is on Terminal-Bench 4.0, where the score rose from 28.3% at preview to 41%. On Cognition’s FrontierCode 2.0 benchmark, Mistral claims 19.7%, which it presents as the best open-weight result. On the Artificial Analysis Coding Agent Index, it reports 48%, and on that firm’s Cyber Index, 51%.

Second, the context window is genuinely large. One million tokens is roughly enough to hold a very large codebase, a long contract set, or months of documentation in a single conversation. For a small team, that changes what you can ask the model to do in one pass — reviewing an entire repository for a security issue, for example, rather than feeding it file by file. It also changes the cost math, because long contexts mean more input tokens per request, and input tokens are what you pay for on the API (more on that below).

Third, the model is positioned for coding and cyber defense, not general chat. Mistral’s pitch leans hard on agentic coding and defensive security work: finding vulnerabilities, reproducing and patching issues, and reasoning through codebases. Independent scoring tempers that pitch. Artificial Analysis’s Intelligence Index v5 rates Large 4 in its high reasoning mode at 26 points — the top Western open-weight model on that index, but fifth among open weights overall, behind Chinese models including DeepSeek V4.1 Flash and GLM-5.3 at 32. In other words: strong at the coding and security tasks Mistral tuned it for, middling as a general-purpose brain. Buy it for the first job, not the second.

Open weights is not the same as open source — read the licence first

This is the part of the story most launch coverage rushes past, and it’s the part that could cost you real money if you get it wrong.

Mistral releases Large 4 under what it calls the Mistral Ecosystem Open-Weight Licence, described as a modified MIT licence. In plain English: you may use the weights, modify them, fine-tune them, and build commercial products on top of them, free of charge. That covers the overwhelming majority of small-team uses — an internal coding assistant, a fine-tuned support bot, a security-review tool you run for your own clients.

The exception is the one that matters if your business model is selling access to the model itself. If you offer Large 4 as a hosted model service — what the licence calls Model-as-a-Service — you need a separate commercial licence from Mistral first. So a startup planning to launch “our own AI API, powered by Large 4” cannot simply download the weights and start charging for access. A consultancy that runs the model internally and sells the results of its work is in a different position from a company selling access to the model. If your plan is anywhere near that line, get the commercial licence question answered in writing before you build on it.

Two more licence-adjacent facts worth knowing. “Open weights” means the trained numbers are downloadable — it does not mean Mistral published the full training data or the complete training process. It published a summary of training content, not a recipe for recreating the model from scratch. And once you download a copy, it is frozen in time: later safety or capability updates Mistral makes to its managed API will not automatically flow into your copy. You own the maintenance burden for anything you self-host, including patching and re-evaluating it.

If you want the full background on how this release was expected to land, our earlier piece on the missed-weights checklist covers the run-up to the release date — and this article replaces its “when will the weights drop?” framing with the real thing. Mistral Large 4 weights missed-date checklist

What it really takes to run Large 4 yourself

Let’s be blunt: the phrase “you can download it” hides a hardware bill that puts self-hosting out of reach for most small teams — and understanding that bill is the whole point of this section.

The full FP8 checkpoint is about 1,060 GB across 54 files. The more compressed NVFP4 version is about 610 GB across 32 files. Serving documentation points to an 8-GPU node for the FP8 weights and a 4-GPU Blackwell node for NVFP4. A 2-bit quantized version has been reported to run on a single Apple Silicon Ultra machine with 512 GB of unified memory — but that’s a heavily compressed variant, and you should expect a capability trade-off for that compression, not a free lunch.

To put that in perspective: Mistral says it trained Large 4 on 3,800 Nvidia Grace Blackwell GPUs in its European data centers. That figure describes training, not what inference requires — nobody is suggesting you buy 3,800 GPUs. But the gap between “a laptop” and “an 8-GPU server node” is the gap between a hobby project and a hosting bill. Renting that class of GPU capacity from a cloud provider typically runs into thousands of dollars a month before you serve a single request, and that’s before engineering time for setup, monitoring, and updates.

There is a middle path that many teams miss: “self-host” does not have to mean “in your office.” It can mean renting GPU capacity from a provider of your choice and running the same open weights there. That still gives you control over your prompts and data — they don’t flow through Mistral’s API — and it lets you shop around on hosting price. It’s still not cheap, but it converts a capital purchase into a monthly bill you can turn off.

My rule of thumb, after working through the numbers: if you don’t already have a GPU hosting setup and an engineer who has run inference servers before, do not start with self-hosting Large 4. Start with the API, measure your real token usage for a month, and only then price out self-hosting against that measured usage. The download will still be there when you’re ready.

The rent-it math: what the API actually costs at real volumes

Mistral reset the API price alongside the weights release: $0.70 per million input tokens, $0.07 per million cached input tokens, and $2.10 per million output tokens. That’s roughly half the preview list prices of $1.36 and $4.18, as documented in the launch-day pricing breakdown. Mistral Large 4 open weights and API pricing The model is also served through Mistral’s own API, OpenRouter, Microsoft Foundry, and Nvidia endpoints, so you can compare host prices rather than taking Mistral’s number as the only option.

Here’s what that price means at volumes a small team might actually hit. These are my worked calculations from the published rates, not Mistral’s marketing examples:

Two things jump out. First, at small-team volumes, the API is almost absurdly cheap compared with self-hosting — $112 a month buys a serious amount of coding assistance, while a single month of rented 8-GPU capacity can cost several times that before engineering time. Second, output tokens cost three times input tokens, so the way you prompt matters: a workflow that asks the model to rewrite entire files on every pass will cost far more than one that asks for targeted diffs.

Then there’s the reasoning-mode lever. Large 4 ships with Low, Medium, and High reasoning modes. Higher modes spend more tokens thinking before answering, which improves results on hard tasks but raises the bill on every request. The practical move is to default routine work — boilerplate, simple refactors, documentation — to Low, reserve High for the genuinely hard problems (a tricky bug, a security review), and measure the difference over a week before locking in a default. If your team treats every request as a High-mode request, you’re paying a premium for thinking you didn’t need.

A note on price confusion you may see elsewhere: some trackers listed sale prices of $0.68 input and $2.09 output against the old $1.36/$4.18 list rates during the preview period. The $0.70/$2.10 figures above are the reset prices reported across October 30 coverage of the weights release. If you see a different number on a pricing page, check the date on the page — preview-era figures are still floating around in older guides, and they’re the ones most likely to mislead you.

Run it vs rent it: the decision table

The pattern here is familiar from every other infrastructure decision: rent while you’re learning, own only when the volume or the privacy requirement justifies it. The open weights change the exit options, not the starting point. Because the weights are downloadable, you’re never locked into Mistral’s API the way you’d be with a closed model — you can move to another host serving the same weights, or bring it in-house later, without rebuilding your prompts and workflows from scratch. That portability is the real gift of this release, even for teams that never download a single file.

What a small team can actually do with it this month

Forget the trillion-parameter headline for a minute. Here’s a practical shortlist, ordered by how quickly each one pays off:

1. Put it on your coding backlog, not your chat window. Large 4’s strongest reported results are in coding and cyber defense. Point it at the unglamorous backlog: dependency updates with changelogs reviewed, test coverage for untested modules, and a first-pass security review of new pull requests. Keep a human in the review loop — the model is an assistant, not an approver — and track how many of its suggestions survive review. That acceptance rate is your real benchmark.

2. Use the 1-million-token context for whole-repo reviews. Feed it an entire small repository plus its issue tracker and ask for a prioritized list of security concerns, with file and line references. Then verify each finding yourself. The value here isn’t that the model is always right; it’s that it reads the whole repo in one pass, which no human reviewer does casually.

3. Fine-tune on your own documentation — but only if you have enough of it. Fine-tuning makes sense when your team has a large, consistent body of internal material (support tickets, runbooks, code conventions) that a generic model keeps getting wrong. It does not make sense for a 50-page handbook. If your corpus is small, prompt engineering with retrieved documents will get you further, faster, and cheaper.

4. Test it against your current model on your own tasks before switching. Run the same 20 real tasks — your actual tickets, not benchmark prompts — through Large 4 and whatever you use today, and score them blind. Vendor benchmarks are a starting point for a shortlist, not a buying decision. If Large 4 doesn’t beat your incumbent on your tasks, the open weights alone aren’t a reason to switch.

5. Don’t build a resale business on it without the commercial licence. I said this above and I’ll say it again because it’s the mistake with the biggest downside: selling hosted access to Large 4 requires Mistral’s separate commercial licence. Selling software or services that use the model is a different thing. When in doubt, ask before you launch, not after.

One more honest note: the earliest community attempts to serve the released weights hit snags — one launch-day report described sessions failing on a chat-template error while other models on the same setup worked fine. That’s normal for a day-one release, and it’s exactly why you should let the serving tooling (vLLM, SGLang) mature for a few weeks before betting a production workflow on a self-hosted setup. The API route doesn’t have that problem.

How Large 4 fits the wider AI price war

Large 4’s price reset doesn’t happen in a vacuum. Across October, frontier API prices have been falling and shifting as labs compete for developer usage, and small teams are the beneficiaries — as long as they keep their workflows portable. AI API price war

The strategic lesson is the same one that applies to every platform decision in this newsletter’s history: the cheapest model today is rarely the cheapest model in six months, and the lab that wins your business today may reprice, pivot, or shut down a product line tomorrow. Open weights are the closest thing to insurance against that. They don’t make the model free — the compute still costs money — but they do mean no single vendor can take the model away from you. In a month when another Google product is being wound down (see our companion piece on Jules), that insurance is worth more than it looks on a pricing page.

Frequently Asked Questions

Is Mistral Large 4 really free to use?

The weights are free to download and use under the Mistral Ecosystem Open-Weight Licence, which permits use, modification, fine-tuning, and commercial products built on the model. The exception is Model-as-a-Service: if you sell hosted access to the model itself, you need a separate commercial licence from Mistral. And “free weights” doesn’t mean “free compute” — running it yourself still costs GPU capacity, while the API charges $0.70 per million input tokens and $2.10 per million output tokens.

Can I run Mistral Large 4 on my own computer?

Almost certainly not at full quality. The FP8 weights are about 1,060 GB and documented serving setups start at an 8-GPU node (4-GPU Blackwell for the NVFP4 version). A heavily compressed 2-bit variant has been reported to run on a single Apple Silicon Ultra machine with 512 GB of memory, but expect a capability trade-off. For most teams, the API or a hosted endpoint is the realistic route.

How good is Large 4, really?

Mistral reports strong coding and security results: 41% on Terminal-Bench 4.0 (up from 28.3% at preview), 19.7% on FrontierCode 2.0, 48% on the Artificial Analysis Coding Agent Index, and 51% on its Cyber Index. But these are Mistral’s own figures, not independently reproduced, and Artificial Analysis’s broader Intelligence Index rates it at 26 — strong for an open-weight model, behind the leading closed models. Test it on your own tasks before switching.

What does the 1-million-token context window actually let me do?

It lets you put very large inputs — an entire small codebase, a long contract set, extensive documentation — into a single request. That’s useful for whole-repo security reviews and large refactoring analysis. Remember that long inputs cost more: every token in that context is billed as an input token on the API, so a million-token prompt at $0.70 per million costs about $0.70 in input alone before the model writes a word.

Should my team self-host or use the API?

Start with the API. At realistic small-team volumes ($11–$112 a month in the table above), it’s far cheaper than renting GPU capacity, and you can measure your real usage before committing. Move toward self-hosting only if you have a data-residency requirement, very high steady volume, or an existing GPU setup — and keep your prompts portable so you can move between hosts serving the same weights.

What happens to my downloaded copy when Mistral updates the model?

Nothing, automatically. Your copy is frozen at the version you downloaded. Updates to Mistral’s managed API won’t flow into it, so you’re responsible for re-downloading, re-testing, and re-deploying when a new checkpoint appears. Factor that maintenance into any self-hosting decision.