Mistral Large 4 is out as an open model, with coding gains since its preview

Mistral AI has released the model files for Mistral Large 4, its largest model, with an announcement on October 29, 2026. The model, which the company calls "Le Chonk", went into preview through the API alone in early October. The files are now on Hugging Face, making it an open model that organizations can download and run on their own servers. According to Mistral, the released version is an improved form of the preview: reinforcement learning (RL) continued over the past three weeks, bringing clear gains in coding and defensive cybersecurity, and the model now does the same work with fewer tokens.

What changed since the preview

GlobalFeed covered the preview in an October 8, 2026 story; at that point the files could not be downloaded and the license had not been announced. According to Mistral's comparisons, the score on Terminal-Bench 4.0, which measures coding agents working in a terminal, rose from 28% in the preview, a figure Mistral measured itself, to 41%, with 30% fewer output tokens. On the B3 benchmark, which tests how well agents resist attacks, the model withstood 97.1% of attacks, up from 93.3% in the preview. On the Finch finance benchmark, the score rose from 67.4% to 73.3%. The model now offers three reasoning levels: low for simple tasks, medium for everyday work, and high, the default, for hard problems. The low level replaces the preview's no-reasoning mode.

Mistral also writes that it monitored cybersecurity-related use in live traffic during the preview. According to the company, the model refused requests to attack live targets, but it saw potential for the model to help with real attacks, so it is changing its guardrails and will keep updating them in the coming months.

Model files and license

According to the model card, Mistral Large 4 is a mixture-of-experts (MoE) model with about 1 trillion parameters. For each token, 4 of 128 experts plus one shared expert are active; that is 49 billion active parameters per token, or 52 billion including the embedding and output layers. The model reads text and images and writes text, supports a 1-million-token context, and works in more than 160 languages, including every official language of the European Union. Mistral trained it from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its own data centers in Europe. The files also include a draft layer called DSpark that speeds up generation without changing the output; Mistral says it measured up to 1.9 times higher throughput with it.

Running the model yourself takes serious hardware. The main files take up about 1.06 terabytes on Hugging Face, and Mistral's vLLM example is set up for a server with eight GPUs. A lighter NVFP4 version is about 610 gigabytes and targets NVIDIA's Blackwell GPUs (B200, B300, GB200, GB300). The model runs with vLLM, SGLang, transformers and llama.cpp. For those who would rather not host it, it is available through the Mistral API, Mistral Studio, OpenRouter, NVIDIA-hosted endpoints and Microsoft Foundry. Organizations can adapt it to their own data with Mistral Forge.

The license is the Mistral Ecosystem license, a modified version of the MIT license. According to the license text, the model can be used, modified, further trained and used commercially free of charge. The one condition concerns companies that run a "Model as a Service" business, offering the model to others through an API or hosted endpoint: they must sign a separate commercial license with Mistral before using it commercially. Internal use, and end-user products where the model is embedded in specific features, fall outside this condition.

Unsloth has added another option by publishing compressed GGUF files of the model. According to Unsloth's guide, the smallest 1-bit file is about 233 gigabytes, fits on a Mac Studio with 256 gigabytes of memory and, in Unsloth's measurement, agrees with the full model on its top next-token prediction about 91% of the time. Unsloth gives no speed figure for this file.

Results shared by Mistral

In coding, Mistral says the model scored 19.7% on Cognition's private FrontierCode 2.0 benchmark, the highest among the open models in the company's chart; the closest, DeepSeek V4.1 Flash, scored 19.3%. On Artificial Analysis's coding agent index, the model scored 48%. GLM-5.3 sits at 49% in the same chart; Mistral calls the two effectively tied and highlights that its model uses 47% fewer tokens per task on average. In blind comparisons run by Surge, graders preferred Mistral Large 4 over GLM-5.3, Kimi K3 and DeepSeek V4.1 Flash; among the models tested, only Claude Opus 5.5 came out ahead. According to the model card, the model scored 26 on Artificial Analysis's overall intelligence index. Mistral presents this as the highest score among Western open models and notes itself that leading Chinese open models score higher on this index.

In cybersecurity, the model scored 51% on Artificial Analysis's cyber index. In Mistral's own chart, MiMo-V2.6-Pro (56%) and DeepSeek V4.1 Flash 1017 (57%) score higher on this index. The company frames its strength as handling defensive work without needless refusals while refusing malicious cyber requests 96.6% of the time on average, which Mistral says is the highest among the open models it measured. Mistral says that since the preview it has worked with leading cybersecurity vendors to test the model in real environments. The announcement includes statements from executives at Booz Allen Hamilton, Cisco, Cloudflare, CrowdStrike, Elastic, Fortinet, Orange Cyberdefense, Rubrik and Zscaler, as well as from ENISA, the European Union's cybersecurity agency. The company has also opened a specialized version built for penetration testing and red-teaming to vetted enterprise partners such as Orange Cyberdefense under contract; it may only be tested on systems the partner owns or is explicitly authorized to test.

The scores and rankings in this story come from Mistral's announcement and model card as of October 29, 2026. Some come from third-party leaderboards such as Artificial Analysis, Cognition and Surge, but Mistral chose which models to compare against and how to present them, and rankings can change as new models are added. For a fuller picture, follow reviews from independent researchers too.

Sources: Mistral AI announcement, Hugging Face (model card), License text, Unsloth guide