Mistral Large 4 Open Weights: A Trillion-Parameter Model You Can Download and Run

On October 29, 2026, Mistral AI released the downloadable weights of Mistral Large 4, nicknamed Le Chonk. That turns Europe's largest open model from a hosted preview into software you can download, inspect, fine-tune, and run on your own servers (Mistral's announcement, retrieved October 30, 2026). The model is a 1.05 trillion parameter, natively multimodal system built for agentic coding and defensive cybersecurity, and its weights are published on Hugging Face under a modified MIT license with one important condition for hosted model services (Hugging Face model card, retrieved October 30, 2026). This article explains what shipped, what the benchmarks show, what the license allows, and what self-hosting takes.

Key Takeaways

• Mistral released the weights of an improved Mistral Large 4 on October 29, 2026, three weeks after its October 6 public preview, following additional reinforcement learning.

• The model totals 1.05 trillion parameters but activates only 49 billion per token, with 52 billion counting embeddings and output layers, and it offers a 1 million token context window.

• Mistral reports 19.7% on FrontierCode 2.0, 48% on the Artificial Analysis Coding Agent Index, and 41% on Terminal-Bench 4.0, plus 51% on the Artificial Analysis Cyber Index.

• The weights use the Mistral Ecosystem Open-Weight License, a modified MIT license. Use, modification, fine-tuning, and commercialization are free, but a Model-as-a-Service business needs a separate commercial license.

• You can download the FP8 checkpoint or the NVFP4 build for Blackwell GPUs, serve it with vLLM or SGLang, or use the hosted API at $0.70 per million input and $2.10 per million output tokens.

Large 4 first appeared on October 6, 2026 as an API-only public preview, with weights promised by the end of October. The October 29 release keeps that promise, but it is not a re-upload of the preview weights. Mistral says the released checkpoint is improved by three more weeks of reinforcement learning and by preview user feedback. That distinction matters for every number here: preview scores describe a replaced checkpoint.

What Is Mistral Large 4?

Mistral Large 4 is Mistral AI's flagship general purpose model, nicknamed Le Chonk for its size. It accepts text and image input, produces text output, and is designed as a coding agent and defensive cybersecurity assistant. Mistral trained it from scratch in Europe on 3,800 NVIDIA Grace Blackwell GPUs in its own data centers, according to the model card.

Mixture of Experts (MoE) is an architecture in which a model holds many specialized sub-networks, called experts, and routes each token to only a few of them. Large 4 holds 128 routed experts, selects 4 per token, plus 1 shared expert. That routing explains the gap between the headline and working numbers: 1.05 trillion parameters in total, but only 49 billion active per token, or 52 billion with embeddings and output layers.

The model also carries a 1.6 billion parameter Pixtral vision encoder for image input. It offers a 1 million token context window, native function calling, and strict JSON schema output. It also carries an embedded DSpark speculative decoding head for serving throughput. It supports more than 160 languages.

What Three More Weeks of Training Changed

Mistral's headline coding claim is a lead among open-weight models. On Cognition's private FrontierCode 2.0 benchmark, which judges whether a maintainer would merge an agent's change, Large 4 scores 19.7% (Mistral's announcement). On the Artificial Analysis Coding Agent Index it scores 48%, effectively tied with GLM-5.3, while using 47% fewer tokens per task on average. On Terminal-Bench 4.0 it improved by 13 percentage points to 41% over the preview, with output tokens down 30%.

Human preference points the same way, with a limit. In Surge Code blind pairwise evaluations, graders preferred Large 4 over GLM-5.3, Kimi K3, and DeepSeek V4.1 Flash. They ranked it behind only Claude Opus 5.5 among the models evaluated. The broader numbers also moved. Finch, for finance spreadsheet work, reached 73.3%, second only to Kimi K3 among open models. BrowseComp, for web browsing work, reached 84.5%, up from 75% for the earlier checkpoint.

Three reasoning modes let developers trade cost against performance. Low suits simple tasks and replaces the old None mode. Medium suits everyday tasks. High is the default for the hardest cases. That control matters: token use, not just raw score, decides what an agentic workflow costs at scale.

One caution applies to all these figures. They are Mistral's own measurements. FrontierCode 2.0 is a private benchmark. Outside researchers cannot rerun it. Mistral says a full technical report is coming soon. Until it arrives, treat the scores as the vendor's claim, not as independently reproduced results.

Built for Defenders, With a Warning Attached

Cybersecurity is the second half of the pitch. Large 4 scores 51% on the Artificial Analysis Cyber Index, which measures discovering, reproducing, validating, and patching a vulnerability without breaking existing functionality (Mistral's announcement). Its built-in guardrails refuse an average of 96.6% of malicious cyber requests across JailbreakBench, StrongREJECT, and AgentHarm prompts, the highest refusal rate Mistral measures among open models. On the B3 Agent Security Benchmark it resists 97.1% of attacks, up from 93.3% in the preview.

Mistral is unusually frank about the other side. The company monitored live cybersecurity traffic. It saw the model refusing black-box attacks against live targets. It still saw a potential for uplift in conducting real attacks. It says it is modifying its guardrails accordingly. A specialized variant is not part of the public download. That variant is for penetration testing, vulnerability assessment, and red-teaming. It goes only to vetted enterprise design partners such as Orange Cyberdefense. Testing is for owned or explicitly authorized systems in controlled environments.

That split is the honest shape of an open-weight security model. The public weights give defenders a frontier tool, but they cannot be recalled once downloaded, and the same capability could help an attacker who already has access. Pierre Stock, Mistral's Vice President of Science, told TechCrunch that open weights give organizations frontier intelligence to defend themselves. That access cannot be cut off by definition because the model is open-weight (TechCrunch, retrieved October 30, 2026).

The License: Open Weight, Not Open Source

Open-weight is the precise term for this release. It means the trained weights are published and downloadable. It does not mean an unrestricted open source license, and it does not promise open training data or training code. Calling Large 4 simply open source, as some launch headlines do, overstates what Mistral granted.

The governing document is the Mistral Ecosystem Open-Weight License, which Mistral describes as a modified version of the MIT license. It permits free use, modification, fine-tuning, and commercialization. It carries a single condition: businesses that operate a Model-as-a-Service offering must obtain a separate commercial license from Mistral before using the model commercially. A Model-as-a-Service business offers third parties access to inference or further training with substantial control over inputs, parameters, and training data.

For most builders the condition never triggers: running the model inside your own product, fine-tuning it on your own data, and deploying it on your own infrastructure are all covered. The line matters if your business resells access to the model itself as a hosted service. Independent coverage notes that Mistral's announcement does not draw a sharp boundary here. A team planning a hosted offering should read the license text first. When in doubt, ask Mistral before building on resold access (independent coverage of the license, retrieved October 30, 2026).

What It Takes to Run It Yourself

Self-hosting starts on Hugging Face, where the weights are published as mistralai/Mistral-Large-4-1T-A52B alongside the NVFP4 quantized checkpoint. Mistral's serving path uses vLLM or SGLang. The NVFP4 build targets Blackwell GPUs such as B200, B300, GB200, and GB300. Those GPUs carry dedicated hardware for that format. The build is meant to cut memory use substantially against the FP8 checkpoint.

That is datacenter-class hardware, so teams should price it before promising a self-hosted rollout. The cheaper route is the hosted 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 price. Microsoft Foundry is the first cloud provider to carry the model. NVIDIA offers an optimized NIM container for serving with vLLM. The model is also available through Mistral Studio, the Mistral API, OpenRouter, and Vibe Code.

If you only want to try it, start with the hosted API. Run your own coding and security tasks. Compare the token bill with the model you use today.

What Independent Coverage Says

TechCrunch frames the release as a sovereignty move in the contest between open-weight developers and closed-model providers. Its coverage stresses that businesses and governments can now download and deploy the system themselves, and it repeats Mistral's claim of performance comparable to GLM-5.3 at 47% fewer tokens per task (TechCrunch). That framing matches Mistral's own announcement, which presents European training and self-hosting as the foundation for sovereign AI.

Coverage also leaves three questions open. First, no independent reproduction of the new scores has appeared. The technical report is still promised, not published. Second, the real cost of self-hosting dwarfs the headline API price. Hardware, power, and engineering time all count. Third, the Model-as-a-Service boundary in the license is untested in practice. None of these cancel the release. Together they mean one thing: test the model on your own tasks before betting a product on Mistral's scoreboard.

Who Should Download It, and Who Should Not

Download it if you run coding agents or defensive security work on infrastructure you control. Download it if you need to fine-tune on proprietary data that cannot leave your network. The reasoning modes, function calling support, and strict JSON schema output are built for those workflows.

Skip the download if you want a small model for a laptop, a hosted service with no operations burden, or a license that lets you resell model access without further permission. Skip it, too, for offensive security testing on systems you do not own: the public weights refuse malicious requests by design, and the capable penetration testing variant is restricted to vetted partners working on authorized systems.

Frequently Asked Questions

Is Mistral Large 4 open source? No, not in the strict sense. It is open-weight: the trained weights are downloadable under the Mistral Ecosystem Open-Weight License, a modified MIT license. Use, modification, fine-tuning, and commercialization are free, but a Model-as-a-Service business needs a separate commercial license from Mistral first.

Is Mistral Large 4 free? The weights are free to download and use under the license terms. Running them is not free: self-hosting needs datacenter-class GPUs, and the hosted API costs $0.70 per million input tokens, $0.07 per million cached input tokens, and $2.10 per million output tokens.

How big is Mistral Large 4, and what hardware does it need? It totals 1.05 trillion parameters, with 49 billion active per token, or 52 billion with embeddings and output layers. It ships as an FP8 checkpoint and an NVFP4 build for Blackwell GPUs, served with vLLM or SGLang. Treat it as datacenter hardware, not a laptop download.

How does Mistral Large 4 compare with GLM-5.3? Mistral reports a tie on the Artificial Analysis Coding Agent Index at 48%, with 47% fewer tokens per task than GLM-5.3, plus leads on FrontierCode 2.0 at 19.7% and Terminal-Bench 4.0 at 41%. Those are Mistral's measurements, and the full technical report is still coming, so reproduce the comparison on your own tasks before deciding.

Can I use Mistral Large 4 for penetration testing? Not with the public weights on systems you do not own. The public model is built for defensive work and refuses malicious cyber requests, with a 96.6% average refusal rate in Mistral's tests. A specialized penetration testing variant exists, but only for vetted enterprise design partners, on owned or explicitly authorized systems in controlled environments.

Our take: the most important number here is not 19.7% or 48%. It is 47%. A model that ties the leading open-weight rival while using 47% fewer tokens per task changes the unit economics of coding agents. For example, a team running an agent loop for an hour pays for every tool call and every retry. Fewer tokens per task means a lower bill for the same work. Mistral pairs that efficiency with a license that keeps hosted resale under its control and an API price cut to about half the preview price. The strategy is clear: give away the weights, keep the hosted business. That is a sound deal for defenders and builders, as long as they read the license line before reselling access.

How this article was checked: we analyzed Mistral's announcement and the Hugging Face model card, then compared them with TechCrunch coverage. This article passed editorial review before publication.

The Bottom Line

Mistral Large 4's open weights are the most consequential European model release of the month because they change who controls the model, not just who can call it. A team that downloads the weights owns its deployment, fine-tuning, and data path, paying in hardware and engineering effort instead of API rent. The license keeps one door closed, reselling the model as a service, and the benchmarks still await independent reproduction. But for coding agents and defensive security work, this is a frontier model you can now test on your own terms. Start with the Hugging Face weights and the Mistral announcement. For context on measuring coding agents, read our Microsoft Scope article, and for the serving side, our DeepSeek DeepInfer article. Our earlier roundup covered the October 6 preview in our This Week in AI article, or read the Indonesian version of this article.