Mistral Large 4 Open Weights: 1T MoE, Modified-MIT License

Introduction

On October 29, 2026, Mistral AI published the weights of Mistral Large 4, the trillion-parameter multimodal mixture-of-experts (MoE) model it nicknames "Le Chonk," on Hugging Face. The model had been in API-only public preview since October 6, with weights promised by the end of the month (our preview article covers that stage). Mistral says the released checkpoint is an improved version, trained with about three weeks of additional reinforcement learning, so preview scores do not carry over to it.

This article covers what actually shipped: the license, the checkpoints, what it takes to run them, and the scores Mistral reports. Every benchmark figure below is vendor-reported unless stated otherwise, and none was independently reproduced for this article.

Feature Overview

Checkpoints. Hugging Face lists two official variants, FP8 and NVFP4, plus GGUF conversions from the llama.cpp team and Unsloth. Sizes below are totals of the file sizes returned by the Hugging Face API on November 2.

Architecture. According to the model card, the model has 128 routed experts with 4 active per token plus one shared expert, about 1T total parameters and 49B active per token (52B including embeddings and output layers). It accepts text and images and outputs text, with a 1M-token context window and three reasoning efforts (low, medium, high, with high as the default). Hugging Face's launch post adds a 72-layer decoder using Multi-head Latent Attention with a DeepSeek Sparse Attention indexer that selects up to 2,048 key positions per query, and a 1.6B Pixtral-style vision encoder. The checkpoint embeds a DSpark speculative-decoding draft head; Mistral says it measured up to 1.9x higher throughput than serving the target model alone, depending on hardware and workload.

License. The weights use the Mistral Ecosystem Open-Weight License, which describes itself as a modified version of MIT. The LICENSE file permits use, copying, modification, distribution, sublicensing and sale, with two practical conditions. First, the copyright notice must travel with copies. Second, if a licensee or its affiliates operates a "Model as a Service" business, they must sign a separate commercial license with Mistral before using the model or derivatives for any commercial purpose. The file defines that term as giving a third party access for inference or post-training, for instance via API or hosted endpoints, in a way that lets them exercise meaningful control over inputs, parameters or training data. Internal use, and use through Mistral's official products, are excluded, as are end-user products with model capabilities solely embedded in specific features.

Vendor-reported benchmarks.

The last row comes from Mistral's own model card, which concedes the aggregate-index gap. Note also that Hugging Face's launch post still quotes preview-era numbers (28.3% on Terminal-Bench 4, 93.3% on B3), so check which version a figure belongs to.

Cybersecurity positioning. Mistral reports 51% on the Artificial Analysis Cyber Index, which covers discovering vulnerabilities in a codebase, reproducing them and patching without breaking existing functionality. It also reports a 96.6% average refusal rate on malicious cyber prompts drawn from JailbreakBench, StrongREJECT and AgentHarm, which it calls the highest it measures among open models. The company adds a caveat: after monitoring live preview traffic, it saw "a potential for uplift to conduct real attacks" and says it is modifying its guardrails. A specialized cybersecurity variant is in preview for vetted enterprise design partners only. Wired reports that Mistral now calls Le Chonk the "world's best" open-weight model for automated coding and cybersecurity; the "leapfrogged China" framing is Wired's headline, not a Mistral benchmark result. Partner endorsements on Mistral's page (Booz Allen, Cisco, CrowdStrike and others) are curated testimonials, not independent tests.

Usability Analysis

Self-hosting is a data-center job. By our arithmetic from the listed file sizes, the FP8 weights alone average about 133 GB per GPU across the eight-GPU tensor-parallel setup in the model card's vLLM example, before any KV cache. The NVFP4 card's SGLang example uses four GPUs, which works out to roughly 150 GB of weights per GPU. Hugging Face's own vLLM example for NVFP4 caps the context at 131,072 tokens rather than the advertised 1M, and its Transformers example launches eight processes with torchrun. The model card itself suggests the Mistral API if local serving is subpar.

Teams that do not want to host have options: the Mistral API at $0.70 input, $0.07 cached input and $2.10 output per million tokens, plus Mistral Studio, OpenRouter, Microsoft Foundry and NVIDIA-hosted endpoints, per Mistral's post. The license also shapes who can build on it. An enterprise running the model internally is exempt from the commercial-license condition, while a company offering hosted endpoints where third parties get meaningful control over inputs, parameters or training data falls under the Model-as-a-Service definition.

Pros and Cons

The strengths are license clarity for most users, an official quantized build, and a stated focus on defensive security. The limits are the hardware bar, a benchmark picture that leans on Mistral's own reporting, and the dual-use question that open cyber-capable weights raise. In the same Wired article, Sophos's Rafe Pilling calls open weights "clearly a dual-use technology."

Outlook

Mistral says a full technical report is coming soon, and that Mistral Large 4 will serve as the foundation for a new generation of specialized models. It also says it will keep updating its cyber guardrails in the coming months. Independent reproductions of the coding and cyber scores, and measurements of how the refusal behavior holds up once weights are fine-tuned, will say more than the launch charts.

Conclusion

Mistral Large 4's open weights are a substantial release: a trillion-parameter model with an official NVFP4 build, day-one runtime support and a license that is permissive unless you run a Model-as-a-Service business. It suits organizations with GPU clusters and data-sovereignty requirements, especially for security and coding agents. Treat the benchmark claims as vendor-reported until third parties confirm them, and read the license before building a hosted product on it.

Editor's Verdict

The open-weight release of Mistral Large 4 earns a solid recommendation for teams that can actually host a trillion-parameter model.

The strongest case for paying attention: a license permissive enough for commercial use, fine-tuning and redistribution, with the Model-as-a-Service condition as the single stated restriction. Reinforcing that, a ready serving path, with official FP8 and NVFP4 checkpoints, a vLLM Docker image, an SGLang cookbook and community GGUF builds — practical value rather than just headline appeal. The broader signal worth registering is straightforward: the license, not the download, is the real gate — internal use is exempt, but offering hosted endpoints to third parties requires a separate commercial license from Mistral. On the other side of the ledger, one constraint is real rather than a marketing footnote: the memory footprint is very large, with the FP8 files totaling about 1.06 TB and even the NVFP4 build near 610 GB. It should factor into any serious decision. Layered on top of that, a benchmark picture that rests mostly on Mistral's own reporting, with the technical report still listed as coming soon — which narrows the set of teams for whom this is an obvious yes.

For organizations with multi-GPU Blackwell or H200-class clusters, data-sovereignty requirements, or security and coding-agent workloads, this is a serious evaluation candidate, not just a curiosity to bookmark. Anyone planning a hosted service on top of it should read the Model-as-a-Service clause first.