Mistral Large 4 Open Weights Are Out: A 1 Trillion Parameter Model for $0.70 Per Million Tokens

Mistral released the open weights for Mistral Large 4 on 29 October 2026, three weeks after the public preview that opened on 6 October. The model is nicknamed Le Chonk in Mistral's own announcement, and Mistral calls it its largest and most capable model to date. It is a natively multimodal mixture of experts model, which means it is built from the start to handle text, images, and other inputs in one model, rather than bolting a vision add on onto a text model later. The headline numbers are large, and they need careful reading: 1 trillion total parameters, 52 billion active parameters including embeddings and output layers, or 49 billion active parameters per token on the more precise split that Mistral's own documentation uses. The practical point is simpler. Only a fraction of the model runs for each token, which is why a trillion parameter model can be served at a price that looks like a mid size model price.

The release matters for a different reason than size alone. Open weights mean you can download the model, run it on your own infrastructure, fine tune it, and inspect how it behaves, subject to the licence. A hosted API alone does not give you that control. A weight release does, and it is why developers, security teams, and European companies that need data to stay under European law have been waiting for this file since the preview. Mistral trained the model from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its own datacentres in Europe, and it serves the preview and the release on that same infrastructure. That European hosting story is part of Mistral's pitch, and it will matter to teams that cannot send client data to a United States cloud for legal or contract reasons.

This article covers what was actually released, what it costs, what the licence allows, what Mistral claims on benchmarks, and the checks to run before you move a real project onto it. Benchmarks in this article are Mistral's own, unless stated otherwise, and independent validation will take time. Do not treat a launch benchmark as a result on your codebase.

What did Mistral actually release on 29 October?

Mistral released an improved checkpoint of Large 4, a second saved set of weights that reflects further reinforcement learning after the preview. Mistral says the reinforcement learning run continued through the preview period, with real world testing alongside government and enterprise users. The weights are published on Hugging Face under the model name Mistral-Large-4-1T-A52B, for deployment with vLLM or SGLang, the two serving frameworks most teams use for open models. The model is also available the same day through Mistral Studio, the Mistral API, OpenRouter, NVIDIA hosted endpoints, and Microsoft Foundry as the first cloud provider, so you do not need to self host to try it.

One detail to check before you download: early coverage on 30 October noted that the main Hugging Face checkpoint page still showed an upcoming release note with an October 31 estimated date, while a separate NVFP4 quantized checkpoint was already available for self hosting. Quantized means the weights have been compressed to use less memory, at some cost to precision. That kind of split is normal on a launch day, when different file formats land at different times, but it means you should confirm the exact checkpoint, format, and licence file on the model page before you start a download, rather than assuming the first file you see is the full precision release. Mistral does not state a memory requirement for self hosting in the sources checked for this article, so this article does not claim a GPU count for a home setup. Check the model card for the format you plan to run, and test with a small deployment before you buy hardware for it.

The model ships with a 1 million token context window, according to Mistral's documentation. That is the amount of text and other input the model can consider in one request. A window that large matters for agentic coding, for security review across a whole codebase, and for finance and legal work where a full deal file or a long contract set must be read together. It also matters for cost, because a large context is only cheap if the pricing and the caching behaviour make rereading that context affordable. Mistral's cached input price, covered below, is the lever that makes a large window practical for repeated work.

Mistral also adds three reasoning modes, Low, Medium, and High, so developers can tune cost against performance. High is the default. That choice matters for budgeting: a task run in High mode will use more tokens, and therefore cost more, than the same task in Low mode. If you are comparing Large 4 with another model, compare the mode as well as the model name, because a High mode result and a Low mode result are not the same product.

What does Large 4 cost, and why is the price about half the preview price?

Mistral's current API pricing for Large 4 is $0.70 per million input tokens, $0.07 per million cached input tokens, and $2.10 per million output tokens. The preview price listed in Mistral's documentation before the release was $1.36 per million input tokens, $0.14 per million cached input tokens, and $4.18 per million output tokens. The new price is therefore about half the preview price on each line: input down about 48.5 percent, output down about 49.8 percent, on the arithmetic from those two published figures. Mistral's announcement states the new price without calling it a cut, and at least one tracker describes it as launch pricing. This article does not state an expiry date for the launch price, because the sources checked do not give a firm end date that can be verified on Mistral's own pages. Budget a long term project on the higher preview rates, and treat the current $0.70 and $2.10 rates as the price today, not as a permanent promise.

To make those rates concrete, use a simple translation. At $0.70 per million input tokens and $2.10 per million output tokens, a job that uses 10 million input tokens and 2 million output tokens costs $7 plus $4.20, which is $11.20. A job that uses 100 million input tokens and 20 million output tokens costs $70 plus $42, which is $112. A job that rereads the same 100 million tokens from cache, rather than as fresh input, costs $7 for that cached input instead of $70, plus whatever output the job produces. That caching saving is why a security review that rereads a codebase many times can stay affordable, while a job that writes long reports on every run will always be output heavy.

Table: Job, illustration on current Mistral rates | Token mix | Cost at $0.70 input, $0.07 cached, $2.10 output. Rows: Small coding task | 1 million fresh input, 200,000 output | $0.70 plus $0.42, total $1.12. Medium agent run | 10 million input, 2 million output | $7 plus $4.20, total $11.20. Cached heavy code review | 100 million cached input, 5 million output | $7 plus $10.50, total $17.50. Large long context job | 100 million fresh input, 20 million output | $70 plus $42, total $112. Same job at preview rates | 100 million fresh input, 20 million output | $136 plus $83.60, total $219.60.

Those figures are illustrations on the published per million rates, not quotes for a real project, because real jobs mix fresh and cached input, and output length varies with the reasoning mode. The useful comparison is the last two rows: the same large job costs about half as much at the release price as it did at the preview price. If you tested Large 4 during the preview and decided it was too expensive, retest it at the release price, and test it in Medium mode as well as High mode, before you rule it out.

What does the licence actually allow, and where is the trap?

Mistral releases Large 4 under the Mistral Ecosystem Open-Weight License, which Mistral describes as a modified version of the MIT licence. For most users, the terms are generous: free use, modification, fine tuning, and commercial use for personal, academic, and general business purposes. You can build a product with it, you can fine tune it on your own data, and you can run it inside your company without paying Mistral a licence fee for the weights themselves. You still pay for API usage if you use Mistral's hosted API, and you still pay for your own compute if you self host, but the weights are not sold per seat.

The one condition that changes the picture is for businesses that operate a Model as a Service offering. If you run Large 4 as a hosted service and resell access to the model itself, Mistral requires a separate commercial licence before you do so. That is the trap for a specific kind of startup: a company that wraps Large 4 in an API and sells that API to others is not covered by the free licence for that use. A company that builds an app, an agent, or a finished product on top of Large 4, where customers buy the product and not access to the raw model, is in a different position, but the line between a finished product and a Model as a Service offering is not defined in the announcement summary, and at least one analysis notes that Mistral does not define where software as a service wrapping ends and Model as a Service begins. Do not guess on that line with client money at stake. If your business model is to resell model access, ask Mistral for the commercial licence terms in writing before you launch. If your business model is a finished product, keep a copy of the licence file from the model page with your project records, so you can show which terms you relied on.

There is a second licence caution. Mistral calls Large 4 open weight, not open source, and that distinction is deliberate. Open weight means the weights are available. Open source, in the strict sense, also implies a licence and a set of freedoms that this modified MIT licence narrows for one class of use. Do not describe Large 4 to clients as fully open source without that qualifier, and do not promise a client that they can resell the model itself unless you have the separate licence. Accuracy on this point is what keeps a proposal honest, and it is also what keeps you out of a licence dispute later.

What do Mistral's benchmarks claim, and what should you test yourself?

Mistral reports three headline benchmark results for the released checkpoint, all as its own results. On FrontierCode 2.0, a coding benchmark from Cognition, Large 4 scores 19.7 percent, which Mistral describes as the top open weight score. On the Artificial Analysis Coding Agent Index, Large 4 scores 48 percent, effectively tied with GLM-5.3, while using 47 percent fewer tokens per task on average, according to Mistral. On Terminal-Bench 4.0, a benchmark for terminal based agent tasks, Large 4 scores 41 percent, up from 28.3 percent for the preview checkpoint, a gain of about 13 points that Mistral attributes to the continued reinforcement learning run. Those are strong claims for an open weight model, and they point to the same strengths: agentic coding, cybersecurity, and enterprise agent work, which is how Mistral positions the model.

Independent context tempers the headline. One review published on 30 October, based on Mistral's documents and outside evaluators, puts Large 4 at 26.02 on Artificial Analysis Intelligence Index version 5.0 in High mode, against 58.41 for Claude Opus 5.5 in Max mode, and notes that Mistral does not say which checkpoint the API serves. That gap is a reminder that a top open weight score is not the same as a top overall score, and that a general intelligence index and a coding agent index measure different things. For a coding or security team, the coding agent numbers matter more than the general index. For a general assistant, the general index matters more. Pick the benchmark that matches your job, and then run your own test.

A practical test plan takes one afternoon. First, run the same three tasks on Large 4 and on your current model: a code fix in a real file from your project, a security review of a short function with a known flaw, and a long document summary with a question that requires reading the whole document. Second, record tokens, cost, and time for each task, in High mode and in Medium mode, so you can see whether Medium mode saves enough to matter without losing quality. Third, check the outputs for the failure modes that matter in your work: invented function names, broken imports, missing edge cases, and patches that look plausible but do not pass tests. A model that passes your three tasks at half the price of your current model is a good switch. A model that only passes the launch benchmarks is not.

Large 4 is worth trying this week, because the weights are out, the API price is about half the preview price, and the context window is large enough for real agent work. It is not worth rebuilding your whole stack on this week, because the checkpoint formats are still settling, the launch price has no verified end date, and the Model as a Service licence line needs a written answer for resellers. Try it on a bounded project, measure it on your own tasks, keep the licence file, and decide from your numbers, not from the trillion parameter headline.

Key takeaways:

• Mistral released Mistral Large 4 open weights on 29 October 2026, three weeks after the 6 October public preview, under the Mistral Ecosystem Open-Weight License.
• The model has 1 trillion total parameters, 52 billion active parameters including embeddings, a 1 million token context window, and Low, Medium, and High reasoning modes, with High as the default.
• API pricing is $0.70 per million input tokens, $0.07 per million cached input tokens, and $2.10 per million output tokens, about half the preview rates of $1.36, $0.14, and $4.18.
• The weights are on Hugging Face for vLLM or SGLang deployment, and the model is also served through Mistral Studio, the Mistral API, OpenRouter, NVIDIA hosted endpoints, and Microsoft Foundry.
• Mistral's own benchmarks claim 19.7 percent on FrontierCode 2.0, 48 percent on the Artificial Analysis Coding Agent Index with 47 percent fewer tokens per task than the tied leader, and 41 percent on Terminal-Bench 4.0, up from 28.3 percent in preview.
• The licence allows free use, modification, fine tuning, and general commercial use, but businesses offering the model as a service need a separate commercial licence from Mistral first.

Related links: Top Use AI Blog, Best open source LLMs 2026, Open source LLM API pricing

Affiliate disclosure: No affiliate link is used in this article. Links to Mistral, Hugging Face, and TopUseAI posts earn no commission.

Sources: Mistral AI, Introducing Mistral Large 4, 6 October 2026; Mistral AI announcement of the open weights release, 29 October 2026, as reported by Unite.AI, Testing Catalog, Pivot News, and DEV Community on 29 to 30 October 2026; Mistral documentation pricing for Mistral Large 4, as checked 30 October 2026; Hugging Face model page mistralai/Mistral-Large-4-1T-A52B.