Mistral Large 4’s weights are now available to download. That sentence sounds simple, but for anyone who follows open models, it closes a story that has been running for most of October. The French lab Mistral first showed Large 4 to the public on 6 October 2026 as an API preview, with the downloadable weights promised for later. On 29 October 2026, the later arrived: the full weights of the finished model, nicknamed “Le Chonk”, were published on Hugging Face, where anyone can download them and run the model on their own servers.
This matters because there is a big difference between being allowed to use a model through someone else’s service and actually holding the model yourself. When you use a hosted service, you send your data to a provider, you accept their rules, and you pay per use. When you hold the weights, you can run the model on hardware you control, fine-tune it on your own data, and build products on top of it without asking a provider for permission each time. That is the promise of open weights, and it is why a release like this gets attention far beyond the usual model-launch news cycle.
But “free to download” does not mean “free to run”, and it does not mean “runs anywhere”. This article explains what Mistral actually released on 29 October, what the licence allows and what it does not, what hardware you realistically need, and how this release compares with the other big model news of the same week. If you only remember one thing, let it be this: the weights are free, but running a trillion-parameter model is a serious infrastructure job, not a weekend laptop project.
What Mistral released on 29 October
Mistral Large 4 is a mixture-of-experts model, which means it is built from many smaller specialist networks, called experts, and only a few of them are active for each piece of text it processes. The headline numbers are striking: roughly 1 trillion parameters in total, with about 49 billion active for each token, or 52 billion if you count the embeddings and output layers. It takes text and images as input and produces text, and it has a context window of 1 million tokens, which means it can work with very long documents in a single session.
The model had been in public preview since 6 October, available through Mistral’s own service. During that preview, the weights were only a promise. In fact, the weights had originally been promised for 27 October, and at an earlier check during that week, the Hugging Face page still showed the release as “Upcoming”, with an estimated date of 31 October and a licence file that returned an error. The 29 October release closes that promise two days ahead of that estimate, and the released checkpoint is an improved version: Mistral says it reflects three more weeks of reinforcement learning and customer testing after the preview.
The official repository on Hugging Face is named Mistral-Large-4-1T-A52B, and it ships in two main variants. The FP8 checkpoint is the main one, used for fine-tuning and for the standard server setup. The NVFP4 checkpoint is a smaller 4-bit format. Community members have also produced GGUF files, the format read by llama.cpp and similar local tools. Day-one support arrived in the main serving frameworks, including vLLM, SGLang, Transformers and llama.cpp, which means developers did not have to wait for tooling to catch up with the release.
Mistral is pitching this model most strongly at two kinds of work: coding and defensive cybersecurity. The company says the model is built to help organisations find and fix software weaknesses, analyse code, and support security research. It also reports training data in more than 160 languages, including every official language of the European Union, which matters for teams working outside English.
Open weight is not the same as open source
One of the most common misunderstandings about this release is the word “open”. Mistral Large 4 is open weight, not open source in the strict sense. The distinction matters, and it is worth getting right before you plan anything around the model.
“Open weight” means the trained model files, the weights, are available for download. You can inspect them, run them, modify them and build on them. “Open source” in the strict software sense would usually also mean the training code and training data are available, under a licence like MIT or Apache 2.0 with no extra conditions. That is not what happened here.
The weights are released under the Mistral Ecosystem Open-Weight License, which Mistral describes as a modified MIT licence. It permits use, modification, fine-tuning and commercial use, including selling products built on the model. But it carries one important condition: if you, or a company connected to you, operate a business that offers the model itself as a service to third parties, a Model-as-a-Service offering, you need a separate commercial licence from Mistral first. Running the model inside your own company, or inside a product where the model is one component rather than the service being sold, is a different situation from selling access to the model itself. If your business plan involves hosting Large 4 and charging other people to use it, read the licence text carefully and talk to a lawyer before you launch. Do not assume that “open weight” means “no strings attached”.
There is one more deliberate limit in this release. Mistral is keeping a more aggressive, penetration-testing-focused version of the model restricted to vetted partners. The public weights are the general model, aimed at coding and defensive security work. The restricted variant is for advanced offensive-security testing under contractual safeguards, and it is not part of this download. That is a safety decision by Mistral, not an oversight, and it is worth knowing so you do not go looking for a version that was never published.
What “free download” really means
Here is the distinction the headline usually leaves out. The weights themselves cost nothing to download. Nobody charges you for the files. But everything around the files costs something, and for a model of this size, that something is large.
First, there is the download itself. The standard FP8 checkpoint is about 1,060 gigabytes, spread across 54 files. That is more than a terabyte of data. You need a fast, stable internet connection and more than a terabyte of free storage just to hold the files, before you have run a single prompt.
Second, there is the hardware to run it. This is a frontier-class model, and it does not run on a phone or an ordinary laptop. Serving guidance published around the release describes an 8-GPU server setup for the FP8 weights, and a 4-GPU Blackwell setup for the NVFP4 variant. Smaller quantised versions exist in the community, and one report notes that 2-bit versions can run on a single high-end Apple Silicon machine with 512 gigabytes of unified memory, but that is a very expensive machine, not a normal consumer computer. The honest summary is this: check the model card for hardware requirements before downloading, because the requirements depend heavily on which variant and which quantisation you choose, and guessing wrong wastes a terabyte of download time.
Third, there is the running cost. Even if you already own the hardware, a server with multiple high-end graphics cards uses a lot of electricity and needs cooling and maintenance. For most individuals and small teams, the practical route to trying Large 4 is not self-hosting at all. It is the hosted API, which Mistral prices at 0.70 dollars per million input tokens and 2.10 dollars per million output tokens, with cached input cheaper. That API price is separate from the weights, and it is how most people will actually use this model in the short term.
So the fair way to describe this release is: free to download, expensive to run yourself, and affordable to try through a hosted service. Anyone who tells you it is simply “free” is leaving out the second half of the sentence.
A crowded week for model releases
Mistral Large 4 did not arrive in a quiet week. Late October 2026 produced three major model stories in four days, and comparing them shows why the word “released” needs careful reading.
OpenAI’s GPT-6.1 Terra arrived on 26 October. Google announced Gemini 4 Carbon on 28 October. Mistral published Large 4’s weights on 29 October. On the surface, that looks like three launches in a row. In reality, they are three different kinds of event. GPT-6.1 Terra is a hosted model you access through a provider. Gemini 4 Carbon was announced, with attention on benchmark claims, but its weights are not something you can download and run yourself. Mistral Large 4 is the one of the three whose weights you can actually download today.
That difference is the real news. Announcements are cheap; downloadable weights are not. A company can announce a model, publish impressive charts, and still keep the model itself behind its own service. Mistral did the opposite sequence: preview first, then the actual files. For developers, researchers and companies that want control over where their data goes, the downloadable release is the more useful event, even if it gets less attention than a flashy announcement.
The table below puts the three events side by side so the difference is easy to see.
| Model | Date in late October 2026 | What actually happened | Can you download the weights? |
|---|---|---|---|
| GPT-6.1 Terra (OpenAI) | 26 October | Hosted model release | No, hosted access only |
| Gemini 4 Carbon (Google) | 28 October | Announcement | No, not a weights release |
| Mistral Large 4 (Mistral) | 29 October | Weights published on Hugging Face | Yes, downloadable now |
There is also a useful contrast closer to home for readers of this site. An earlier article covered Kimi K3.1 Pro, whose weights were promised by 4 November 2026 and had still not shipped when this batch’s research was compiled. That is a promise with a deadline still in the future. Mistral Large 4 is the opposite case: a promise that has already been kept. The two stories together show both sides of open-weight news, and they are a good reminder to check whether weights are actually out before planning a project around them, rather than trusting a promised date.
What the benchmarks do and do not prove
Mistral’s announcement comes with strong performance claims, and it is important to handle them honestly. The company says Large 4 scores 48 percent on the Artificial Analysis Coding Agent Index, effectively tied with the leading open model there while using 47 percent fewer tokens per task, and 51 percent on the Artificial Analysis Cyber Index. It also reports that the model refused an average of 96.6 percent of malicious requests across several safety tests, and that its score on Terminal-Bench 4.0 rose from 28 percent at preview to 41 percent in the released checkpoint.
Those numbers are Mistral’s own claims, published by the company that built the model. They are not independent test results, and they have not been reproduced by outside testers at the time of writing. Some outside reporting has also noted that on the aggregate Artificial Analysis Intelligence Index, Large 4 scores around 26, which places it fifth among open-weight models overall, behind several Chinese models, even while it leads among Western open-weight models. None of that makes the model bad; it makes the marketing incomplete. A single benchmark never tells the whole story, and a company’s own charts are the most flattering version of that story.
The practical advice is simple. If you are choosing a model for a real job, test it on your own tasks with your own data. A coding benchmark does not guarantee good results on your codebase, and a cybersecurity benchmark does not guarantee good results on your systems. The advantage of downloadable weights is exactly that you can run those tests yourself, on your own hardware, without sending sensitive code or data to a provider. Use that advantage instead of trusting a chart.
Who should care, and who should wait
This release matters most to a specific set of people. If you run a software company, a security team, a research lab or a government agency, and you need a capable model that runs inside your own infrastructure, Large 4 is now a genuine option to evaluate. The combination of downloadable weights, a permissive licence for internal and product use, and a focus on coding and defensive security makes it a serious candidate for teams that cannot or will not send their data to a hosted service.
It also matters to developers building products in Europe. Mistral is a France-based lab, and the model was trained in European data centres. For organisations with data-residency requirements, or for teams that prefer a European provider, that is a real consideration, not just a point of pride.
Who should wait? Most individual users. If you simply want a capable assistant for everyday writing, coding help or questions, the hosted API or any existing hosted service will serve you better, faster and cheaper than downloading a terabyte of weights and renting server hardware. Self-hosting is a project, not a download, and it only makes sense when you have a specific reason to control the infrastructure: privacy, compliance, customisation, or scale.
Students and hobbyists fall somewhere in between. If you want to learn how large models are served, Large 4 is a fascinating case study, but it is not a beginner project. Start with the hosted API, or with a much smaller open model that genuinely runs on consumer hardware, and treat Large 4 as something to study and evaluate rather than something to run on your own machine this weekend.
How to get started safely
If you do decide to try Large 4, here is a sensible order of steps. First, read the model card on Hugging Face before downloading anything. The card lists the variants, the licence, the hardware guidance and the serving examples, and it is the authoritative source for the details that change most often. Second, decide which route fits you: hosted API for trying it out, or self-hosting for a real deployment. Do not download a terabyte of files before making that decision.
Third, if you self-host, start with the serving framework you already know, whether that is vLLM, SGLang, Transformers or llama.cpp, and follow the official deployment guide for the variant you chose. Early user reports after the release included at least one serving failure caused by a chat-template configuration issue, which is a reminder that day-one setups for a model this large still need care and testing. Fourth, test on non-sensitive data first. Run your evaluation, check the outputs, and only then point the model at real code or real security work.
Finally, keep the licence in mind from the start. If your plan ever grows into offering the model itself as a hosted service to others, you will need that separate commercial licence from Mistral. Building first and checking the licence later is an expensive way to learn about licensing.
What to watch next
Several things are worth watching in the weeks after this release. The first is independent testing: as more teams run the released weights, outside results will either confirm or qualify Mistral’s claims, and those results matter more than launch-week charts. The second is the technical report, which Mistral has said is coming but has not yet published with a date. A full technical report will answer questions about training and evaluation that the announcement leaves open.
The third is the wider open-weight race. Large 4’s release puts pressure on other labs to ship weights, not just announcements, and the promised Kimi K3.1 Pro weights, due by 4 November, are the next date to watch. If those ship on time, late 2026 will be remembered as the period when open weights caught up with hosted announcements. If they slip, Mistral’s on-time delivery will look even better by comparison.
The fourth is pricing. Mistral cut its hosted API price alongside the weights release, roughly halving it from the preview period’s list price. Whether that price holds, and whether other providers follow, will shape how affordable frontier-class models are for smaller teams.
For now, the headline stands on its own: a trillion-parameter model from a European lab, with weights you can actually download, under a licence that permits real commercial use with one clear condition. That is a meaningful step for open models, even if running it yourself remains a job for servers rather than laptops.
The release details and dates in this article follow reporting by The Model Press and Mistral’s announcement.
FAQs
Is Mistral Large 4 really free?
The weights are free to download, but running the model is not free. You need large storage for the roughly 1,060-gigabyte download, powerful server hardware to run it, and electricity and maintenance to keep that hardware going. For most people, the hosted API, priced at 0.70 dollars per million input tokens and 2.10 dollars per million output tokens, is the affordable way to try it.
Can I run Mistral Large 4 on my phone or laptop?
No, not in any practical sense. This is a frontier-class, trillion-parameter model. Serving guidance describes multi-GPU server setups, and even the smallest community versions need very high-end hardware. Check the model card for hardware requirements before downloading, and do not expect it to work on an ordinary laptop or phone.
What does the licence actually allow?
The Mistral Ecosystem Open-Weight License, a modified MIT licence, allows use, modification, fine-tuning and commercial use, including selling products built on the model. The one major condition is that businesses offering the model itself as a service to third parties need a separate commercial licence from Mistral. Internal use and product use are not the same as selling access to the model itself.
How is this different from the Kimi K3.1 Pro weights story?
Kimi K3.1 Pro’s weights were promised by 4 November 2026 and had not shipped when this batch’s research was compiled, so that story was still a promise. Mistral Large 4’s weights were published on 29 October 2026 and are downloadable now on Hugging Face. One is a deadline to watch; the other is a release you can already use.