Mistral Large 4 Open Weights: Le Chonk Goes Free

Updated: October 2026 · 6 min read · By the CreatorSprout team

Launch coverage based on Mistral's own October 29, 2026 announcement and independent launch reporting — we have not tested Mistral Large 4 hands-on. Mistral's own claims and independently reported details are kept separate below. No affiliate links in this article.

The timing is the story. Large 4 entered public preview on October 6, 2026, with Mistral promising the weights would follow. They did, three weeks later — and the weights are not the same model that previewed. Mistral says three extra weeks of reinforcement learning plus customer feedback lifted its coding and defensive-cybersecurity results substantially, and it used the release to re-price the API at the same time. In a month that already brought OpenAI's GPT-6.1 Terra, Google's Gemini 4 Carbon and Google's Genie 4, the open-weight frontier just got its biggest new download.

What Mistral Large 4 is

Mistral Large 4 is Mistral AI's flagship model: a natively multimodal mixture-of-experts model with roughly 1 trillion total parameters, about 52 billion of them active per token, and — according to launch reporting — a 1-million-token context window. It takes text and images as input and produces text and code. Mistral trained it from scratch on European infrastructure, and that origin story is the point of the release: the company pitches Large 4 as a foundation for "sovereign AI," where governments and enterprises can own, inspect and customize the model instead of depending on an outside provider's API.

The model also ships with three reasoning modes, new with the weights release: Low for simple tasks, Medium for everyday work, and High (the default) for the hardest cases needing precision and thorough reasoning. That is a cost control, not just a quality dial — a simpler mode on a routine task burns fewer tokens, and Mistral's efficiency claims below rest heavily on it.

The coding claims

Mistral's announcement leads with coding, and the numbers — all Mistral's own, not yet independently reproduced — are:

FrontierCode 2.0: 19.7%. Cognition's private benchmark scores end-to-end code quality and whether a repository maintainer would merge the agent's change. Mistral says Large 4 leads open-weight models here.
Terminal-Bench 4.0: 41%, up from 28% at preview. A 13-point jump in three weeks, with output tokens down 30%, which Mistral presents as the open-weight lead.
Artificial Analysis Coding Agent Index: 48%. Effectively tied with GLM-5.3, Mistral says, while using 47% fewer tokens per task on average — the efficiency half of the claim matters as much as the score, because token count is what you pay for.
Human preference: in Surge Code's blind pairwise evaluations, Mistral says Large 4 was preferred over leading open-weight models including GLM-5.3, Kimi K3 and DeepSeek V4.1 Flash, ranking behind only Claude Opus 5.5 among the models evaluated.

Mistral also shows the model working without human help: building a live 3D map of roughly 19,000 satellites and debris pieces from real tracking data, debugging it in a browser, and fixing what an independent reviewer flagged until all requirements passed. That is a demo, not a benchmark — but it is the kind of end-to-end agentic task the benchmarks above are trying to measure.

The cybersecurity claims

The second half of the pitch is defensive security — finding vulnerabilities, reproducing them, and patching without breaking existing functionality. Mistral reports 51% on the Artificial Analysis Cyber Index for that full lifecycle, and says the model refuses 96.6% of malicious cyber requests on average across JailbreakBench, StrongREJECT and AgentHarm prompts, the highest refusal rate it measures among open models. Agent-security resistance also improved: 97.1% on the B3 Agent Security Benchmark, up from 93.3% at preview.

Mistral is unusually candid about the risk here, and that candor is worth quoting in spirit: monitoring live traffic, it observed the model refusing black-box attacks against live targets, but it still sees potential for the model to make real attacks easier, and says it is tightening guardrails over the coming months. A specialized penetration-testing variant exists, but only for vetted enterprise design partners such as Orange Cyberdefense, under contractual safeguards and restricted to owned or explicitly authorized systems. That variant is not the public download.

For context, this is the same week Google's Gemini 4 Carbon — also aimed squarely at cyber defense — stayed locked behind a vetted-defender program. Mistral's answer is the opposite strategy: publish the weights and let defenders run the model inside their own walls.

The license — read it before you ship

Large 4 ships under the Mistral Ecosystem Open-Weight license, which Mistral describes as a modified MIT license. It is free to use, modify, fine-tune and commercialize — with a single condition that decides whether your plan works:

Internal use, products and fine-tunes: free, no separate license needed.
Model-as-a-Service: if your business sells access to the model itself as a hosted service, you must obtain a separate commercial license from Mistral before using it commercially.

That condition is the difference between "open weights" and "no strings attached." If you are building an app that happens to use Large 4 behind the scenes, the reporting suggests you are fine; if your product is hosted Large 4 access, read the license text on the model card before you launch. Mistral has also said a full technical report is still coming, so some details — including the exact training mixture — remain unconfirmed.

Pricing and access

API pricing from October 29, 2026: $0.70 per million input tokens, $0.07 per million cached input tokens, and $2.10 per million output tokens — roughly half the preview rates, according to launch reporting.
Where to get it: Mistral Studio, the Mistral API, OpenRouter, and Microsoft Foundry for managed deployment; the open weights are on Hugging Face for self-hosting with vLLM or SGLang; NVIDIA offers hosted endpoints and a NIM container. You can also try it in Mistral's Vibe Code by giving it a feature to build or a bug to fix in your own repository.
Customization: Mistral Forge lets enterprises tailor the model with proprietary data, policies and detection logic, and optimize it into smaller task-specific variants.

The hardware reality

Free weights are not free compute. Launch reporting puts the full model at roughly 1.1 TB of GPU memory, or about 610 GB for a compressed NVFP4 version, with the standard build needing eight high-end data-center GPUs. That is why the API price cut matters more than the download for most readers: at $0.70/$2.10, the hosted route is the way almost everyone will actually use this model, and the download is mainly valuable to teams with data-center hardware, or to researchers and enterprises that need the model inside their own walls for sovereignty or compliance reasons. Smaller, distilled variants are the usual next step for open-weight flagships — Mistral says Large 4 will serve as the foundation for a new generation of specialized models — but none has been announced yet.

What it means for creators and developers

For most creators, Large 4 changes nothing you can click and use today that you could not do with a hosted API last week — but it changes the economics and the control story underneath. If you build coding agents, the 47%-fewer-tokens claim (if it holds up) is the number that hits your bill, and the three reasoning modes give you a lever to trade quality against cost per task. If you work in security, the defensive-cyber profile plus self-hosting is a combination no closed frontier model offers: the model can live next to your code and your vulnerability data without either leaving your network.

For everyone else, the practical move is the same as with every open-weight release: wait for independent runs. The benchmark table that matters — Artificial Analysis' overall Intelligence Index — rates Large 4 at 26, fifth among open weights overall and behind Chinese models at 32, even as Mistral leads on the coding-specific indexes it highlights. Both facts can be true at once: best open-weight coder by Mistral's chosen measures, mid-pack general model by an independent aggregate. Judge it on your task, not on either headline alone.

The honest caveats

We haven't tested it. Everything here comes from Mistral's announcement and launch reporting, not our own runs, and the full technical report is still to come.
Most headline numbers are Mistral's own. FrontierCode, Terminal-Bench and the Cyber Index figures are vendor-reported; the independent Artificial Analysis aggregate (26) is notably less flattering than the coding indexes Mistral leads with.
The preview comparison excludes a rival. Mistral notes its comparisons exclude Moonshot's Kimi K3.1 Pro, whose weights were not yet public at release — a model that tops Terminal-Bench 4.0 at 67% in the reporting cited by launch coverage.
Free download, expensive run. Data-center-scale hardware requirements mean the weights are practical for enterprises and researchers, not for a laptop.
The license has a real condition. Model-as-a-Service businesses need a separate commercial license — check the model card before building a hosted offering on it.
Dual-use is acknowledged by Mistral itself. The company says it sees potential for real-attack uplift and is tightening guardrails; the penetration-testing variant stays vetted-only.

Verdict

Mistral Large 4's open-weights release is a genuinely major launch — the biggest open model Europe has shipped, at half the preview price, with a coding story that independent testers now get to verify for themselves. That last clause is the point of open weights: unlike a closed launch, every claim in this article can be checked by anyone with the hardware. If the coding efficiency holds up, Large 4 becomes the default open choice for agentic coding; if the independent aggregate proves closer to the truth, it is still a strong, sovereign, self-hostable option — just not the open frontier's outright leader. Either way, download it if you have the infrastructure, use the API if you don't, and read the license before you resell it.

FAQ

What is Mistral Large 4 (Le Chonk)?

Mistral AI's flagship model, nicknamed Le Chonk: a natively multimodal mixture-of-experts model with roughly 1 trillion total parameters and about 52 billion active per token. It entered public preview as an API-only model on October 6, 2026, and Mistral released its full open weights on October 29, 2026.

Is Mistral Large 4 really free?

The weights are free to download, use, modify, fine-tune and commercialize under Mistral's Ecosystem Open-Weight license, a modified MIT license — with one condition: a business that operates a Model-as-a-Service offering must obtain a separate commercial license from Mistral first. Running it yourself still needs serious GPU hardware.

How much does the Mistral Large 4 API cost?

From October 29, 2026, Mistral prices the API at $0.70 per million input tokens, $0.07 per million cached input tokens, and $2.10 per million output tokens — roughly half the preview pricing, according to launch reporting.

Can I run Mistral Large 4 on my own computer?

Almost certainly not on a normal desktop. Launch reporting puts the full model at about 1.1 TB of GPU memory, or about 610 GB for a compressed NVFP4 version, with the standard build needing eight high-end data-center GPUs. For most people, the practical routes are Mistral's API, Mistral Studio, OpenRouter, or a managed deployment such as Microsoft Foundry.

How good is Mistral Large 4, really?

Mistral claims the open-weight lead on coding (FrontierCode 2.0 at 19.7%, Terminal-Bench 4.0 at 41%) and a 48% score on the Artificial Analysis Coding Agent Index, tied with GLM-5.3 while using 47% fewer tokens. The independent counterweight: Artificial Analysis' overall Intelligence Index rates it at 26, behind Chinese open models such as DeepSeek V4.1 Flash and GLM-5.3 at 32. Treat the coding claims as Mistral's own until independent testers reproduce them.