Mistral Releases Open Weights for Its Trillion-Parameter Large 4
Mistral published open weights for Large 4, a trillion-parameter model it is positioning for cybersecurity and enterprise work, on October 29, 2026. The French AI lab released the model, nicknamed "Le Chonk," on Hugging Face and through its own API, making it one of the largest open-weight models any lab has shipped so far.
⚡ Quick facts
- Release date: Mistral published Large 4's weights on October 29, 2026
- Nickname: "Le Chonk," a reference to the model's size
- Size: About 1 trillion total parameters with 52 billion active per request, per its Hugging Face listing
- Focus: Cybersecurity and enterprise deployments, alongside general-purpose use
- Access: Hugging Face (including a lower-cost NVFP4 version) and Mistral's API, with on-premises deployment possible since the weights are public
What Mistral actually shipped
Large 4 is a mixture-of-experts model: its Hugging Face repository, listed under Mistral's own organization as "Mistral-Large-4-1T-A52B," shows roughly 1 trillion parameters in total with about 52 billion active for any given request, the part of the model that design actually uses at once. Mistral also released a quantized "NVFP4" version meant to run more cheaply, alongside the full model. Both went up on Hugging Face at the end of October, and Large 4 is also reachable through Mistral's own API for developers who prefer not to host it themselves.
Why Mistral calls it "Le Chonk"
Mistral built a small set of playful Hugging Face demo pages around the nickname "Le Chonk," including ones that translate memes and converse in more than 100 languages, a lighthearted way of showing off the model's scale and multilingual range. CEO Arthur Mensch framed the release in less playful terms, saying AI "should be a tool for empowerment, not dominance," a line aimed at the broader debate over whether the most capable models should stay locked inside a handful of companies.
The cybersecurity pitch, in the company's own words
Mistral is pitching Large 4 specifically at security teams. Pierre Stock, the company's VP of Science, argued that an open-weight model has an advantage here because it "can't be cut off by definition because the model is open-weighted," meaning a company using it for security monitoring can't lose access the way it could with a closed, hosted model. Stock also said the model is built to help organizations "identify, coordinate, and respond in real time to major security incidents," and that it can be deployed entirely on a company's own infrastructure. He added that Mistral isn't trying to overstate the danger of AI-driven attacks: "We don't want to fearmonger people."
What's confirmed, and what's still murky
The release date, the model's name, its presence on Hugging Face and its roughly 1-trillion/52-billion parameter split are confirmed by Mistral's own model listing. What isn't yet clear is the exact license. Mistral's earlier flagship, Large 3, shipped under the permissive Apache 2.0 license, but at the time of writing Large 4's Hugging Face page did not display a specific license tag, and outside reporting on the release has described it only as a custom license rather than naming Apache 2.0 or another standard terms. Pricing for Large 4 through Mistral's API also hadn't been published in the launch coverage reviewed for this article.
Benchmarks: Mistral's charts versus independent runs
Mistral's own launch materials show Large 4 topping a legal-agent benchmark built by Harvey, ahead of rivals including Kimi K3 and GPT-6 Astra, and claim leads on a cybersecurity-focused "Cyber Index" and on token efficiency against GLM-5.3. Those are the company's figures from its own charts. The independent benchmark site Vals AI ran its own evaluation of Large 4 the same week and found a more mixed picture: on its broad Vals Index, Large 4 ranked 32nd out of 44 models tracked, with a legal-specific score noticeably behind rivals like DeepSeek V4 Pro and Hy4 Preview, even as it held up better on finance-focused tasks. Readers should treat Mistral's comparison charts as marketing claims and the Vals AI numbers as the closest thing to an independent check available so far.
Who this matters to
The audience here is mostly security teams and developers who want a capable model they can run and inspect themselves rather than call through a closed API, plus enterprises that specifically want AI tooling they can keep running even if a vendor relationship sours. For casual chatbot users, Large 4 changes little day to day since it isn't a consumer product. For companies already evaluating open-weight alternatives to closed models from OpenAI, Google and Anthropic, Large 4's size and Mistral's cybersecurity framing make it one of the more notable entries so far in 2026.
Frequently asked questions
What is Mistral Large 4?
Mistral Large 4 is a mixture-of-experts language model from the French AI lab Mistral, with roughly 1 trillion total parameters and 52 billion active parameters per request, according to its Hugging Face model listing. Mistral released its weights publicly on October 29, 2026, and is positioning it for cybersecurity and enterprise use.
Why did Mistral nickname it "Le Chonk"?
Mistral gave Large 4 the nickname "Le Chonk" as a playful reference to its size, 1 trillion total parameters, and built a set of lighthearted Hugging Face demo Spaces around the name, including ones that translate memes and speak more than 100 languages.
Is Mistral Large 4 open source?
Mistral published Large 4's weights openly on Hugging Face, which lets anyone download and run the model, but at the time of writing the model's Hugging Face listing did not display a specific open-source license tag, and reporting on the release has not confirmed whether it carries the same Apache 2.0 license as the earlier Mistral Large 3.
How does Mistral Large 4 perform on benchmarks?
Mistral's own launch materials show Large 4 leading rivals including Kimi K3 and GPT-6 Astra on a legal-agent benchmark from Harvey. Independent runs by the benchmark site Vals AI put Large 4 around 48% overall, ranked 32nd of 44 models it tracks, with more mixed results than Mistral's own charts, particularly on legal tasks.
Where can I use Mistral Large 4?
Mistral Large 4's weights are available on Hugging Face under the mistralai organization, including a quantized NVFP4 version for cheaper inference, and the model is also accessible through Mistral's own API. Because the weights are public, companies can also deploy it on their own infrastructure.
Comments