Mistral AI released the weights of its largest model, Mistral Large 4, nicknamed Le Chonk, on October 29, 2026. The model is a mixture-of-experts system with roughly 1 trillion total parameters and 49 billion active per token, trained from scratch on 3,800 Nvidia Grace Blackwell GPUs in Mistral's own European data centers. The company positions it as the strongest open-weight model built in the US or Europe, with performance competitive with the best open models worldwide.
The release landed three weeks after a public preview on October 6. Mistral said the final version benefited from three more weeks of reinforcement learning, with gains that show, in its words, "no signs of saturation." The weights ship on Hugging Face in FP8 and NVFP4 checkpoints, under the Mistral Ecosystem Open-Weight License, a modified MIT license. Internal use is free, but a company that sells the model to third parties as a hosted service must sign a separate commercial license.
This matters because open-weight models set the floor on price for AI services. Mistral prices the preview at $0.68 per million input tokens, $0.07 cached, and $2.09 output. That sits a few cents above DeepSeek V4 Pro 0813's OpenRouter listing and well below Mistral's own Medium 3.5 at $1.50 input and $7.50 output. For teams that must run models on their own hardware, or keep data inside Europe, Large 4 is the first trillion-parameter option from a Western lab.
What is in the model
Large 4 is natively multimodal. A 1.6 billion parameter vision encoder handles text and image input, and the context window runs to 1 million tokens. Reasoning effort is set per request to low, medium, or high, with high as the default. An embedded DSpark speculative-decoding draft head, measured at up to 1.9x higher throughput, is built in. The model runs on vLLM, SGLang, and llama.cpp.
Mistral trained the model on data spanning more than 160 languages, including every official language of the European Union. The company says it worked with enterprises in finance, engineering, manufacturing, logistics, pharmaceuticals, science, shipping, and the public sector during training. The same training, customization, and reinforcement learning environment is offered to customers through Mistral Forge.
A practical caveat: a mixture-of-experts model still has to hold all of its weights in memory to serve. Forty-nine billion active parameters cut compute per token, not the hardware needed to load roughly a trillion parameters. Teams planning self-hosted deployments should size memory accordingly.
Where it scores well
On Mistral's numbers, Large 4 posts 61.7% on DeepSWE v1.1, 59.4% on SWE-Atlas-QnA, and 28.3% on Terminal-Bench 4 in the preview, with the weights release pushing Terminal-Bench 4.0 to 41%, up 13 points. Its Coding Agent Index of 49.8% places it ahead of DeepSeek V4 Pro 0813 and Qwen3.8 Max. On AutomationBench, 657 business workflows across Gmail, Google Sheets, Slack, and Salesforce, it scores 59.9%, ahead of Kimi K3, MiMo-V2.6-Pro, and DeepSeek V4 Pro.
Mistral's headline claim is security. On the Artificial Analysis Cyber Index, Large 4 ranks among the top five models globally and leads open-weight models developed outside China by a wide margin. On a test that asks a model to reproduce a real vulnerability in open-source software and patch it, Large 4 scores 82%, the highest of any model. It also solves 93% of the 40 challenges in Cybench, drawn from security competitions.
That 82% against near zero is a refusal gap as much as a skill comparison. Mistral says several leading closed models, including Claude Opus 5.5 and GPT-6 Astra, score near zero on the same test because they refuse to perform the task. Defending software often starts with proving a flaw is real, work that safety filters in closed models can block. Mistral's pitch is a model that does defensive work and turns down attacks, run on private cloud or on premises under the customer's own rules.
On the safety side, Large 4 resists 97.1% of attacks on the B3 Agent Security Benchmark and posts a 96.6% average refusal rate on malicious cyber prompts from JailbreakBench, StrongREJECT, and AgentHarm. On the KORA Benchmark it scores 1.691 out of 2, the highest Mistral has measured among open models.
Where the numbers need scrutiny
Every benchmark figure in Mistral's announcement is Mistral's own, or cited from Artificial Analysis, vals.ai, or Surge AI. Independent reruns on the weights are only now possible, and several of those evaluators' own pages were not public when the preview launched.
On DeepSWE v1.1, a table published by Reflection AI the day before Large 4's preview puts Kimi K3 at 68.0 and DeepSeek V4.1 Flash at 74.2, both above Large 4's 61.7. Large 4 lands level with GLM 5.3 at 61.0. On AutomationBench, Large 4 posts the highest number in that set, but Reflection labels its column the public split while Mistral describes 657 workflows, so the comparison is indicative at best.
FrontierCode 2.0, a private benchmark from Cognition of whether a repository maintainer would merge an agent's change, puts Large 4 at 19.7%, a 0.4 point lead over DeepSeek V4.1 Flash at 19.3%. Outsiders cannot rerun that test. On the Artificial Analysis Intelligence Index v5, Mistral cites a score of 26, which it calls the highest among Western open-weight models, while acknowledging that leading Chinese open models score higher.
Mistral's own wording holds up on inspection: the strongest open model from the US or Europe by its numbers, competitive rather than leading worldwide.
The sovereignty angle
Mistral trained Large 4 from scratch on 3,800 Nvidia Grace Blackwell GPUs in its own European data centers and serves the preview on the same hardware. A European deployment runs end-to-end under European law, independent of other digital service providers. The company says this matters in cybersecurity, where provider-level refusals can block legitimate vulnerability research and incident response, and where losing access to a capability mid-incident can itself become a critical security risk.
The timing tracks Mistral's broader push. The company raised 3 billion euros in a Series D at a valuation above 21 billion euros, the largest equity round ever raised by a European technology company. It opened a German hub in Munich to advance industrial AI and partnered with Mozilla to bring open, private, multilingual AI to Firefox's Smart Window. The company frames Large 4 as Europe's open-weight answer to the Chinese labs.
Pricing and the open-weight question
Mistral prices Large 4 at $0.68 input, $0.07 cached, and $2.09 output per million tokens on its Standard tier. The model page also shows $1.36 input and $4.18 output, exactly double, and the pages Mistral published do not say whether the lower figure is a preview rate. Large 4 was not yet listed on OpenRouter when first checked, so hosted prices from third parties remain an open question.
The license terms changed between announcement and release. The October 6 model page said only Open. The October 29 weights shipped under the Mistral Ecosystem Open-Weight License, which the card calls a modified version of the MIT license. It allows use, modification, and resale, with one condition: a licensee that operates a model as a service business must sign a separate commercial license before commercial use. Internal use and use through Mistral's own products are exempt.
What to watch
Hosted prices for the open weights are the next data point. Once anyone can serve the model, OpenRouter and cloud listings will show whether Mistral's $0.68 and $2.09 hold. Architecture and post-training details, which Mistral promised with the weights, are still pending. Outside evaluations on Terminal-Bench, DeepSWE, and the Artificial Analysis pages behind the coding and cyber numbers will settle how much of Mistral's claim survives independent testing.
Large 4 is also the foundation for a new generation of specialized Mistral models. The reinforcement learning run behind the preview is still in flight, and Mistral says the model shows no signs of saturation. More compute means more training, and the company expects large and rapid improvements in the weeks and months ahead.
For teams that must keep data in Europe, Large 4 is Mistral's largest open-weight model yet, and it runs a European deployment end to end under European law. For everyone else it adds another strong open option beside DeepSeek, Kimi, GLM, and Qwen, at a price close to DeepSeek V4 Pro. The weights are on Hugging Face now; the independent test results are not.
See more AI coverage at our AI category. Read Mistral's announcement for the full benchmark tables.