Mistral Large 4 Released on Hugging Face: Inside the 1-Trillion-Parameter Model
機械翻訳 / Machine-translated
A
AI Friends
5日前・15分で読める
機械翻訳 / Machine-translated
@aifriends
AI Friends（https://aifriends.jp）のクロスポスト公式アカウント。AIツールの紹介・使い方・できることを、中学生でもわかるやさしい日本語で届けます。
Wouldn't you be surprised to hear that "a 1-trillion-parameter AI is now available for anyone to download"? French AI company Mistral AI has released a new model. This article walks through what's inside, pricing, and key caveats.
Mistral Large 4 (nicknamed "Le Chonk") is a large AI model developed by French company Mistral AI. An API preview launched on October 6, 2026, and the weights (the data that make up the AI's brain) were published on Hugging Face on October 29.
Hugging Face describes the model in its official blog as "a MoE (Mixture-of-Experts) model that takes text and images as input and produces text output." MoE is a mechanism that selects only a small number of specialized sub-networks from a large pool and activates them as needed.
At the time of the preview launch, Mistral had stated it would release the weights by the end of October, citing additional safety reviews as the reason. That promise was fulfilled in approximately three weeks.
The total parameter count (a measure of the model's size) is approximately 1 trillion. However, when processing a single token, only about 49 billion parameters are active. The documentation also mentions 52 billion when including input and output layers.
According to Hugging Face's explanation, each layer has 128 experts, 4 of which are selected per token, along with 1 shared expert. It works like opening only the shelves you need in a vast library.
The context length (the amount of text the model can read at once) is 1 million tokens — enough to feed it several books' worth of material at once. However, third-party evaluations put this figure at around 520,000 tokens, and numbers vary depending on the source.
Images are processed by a Pixtral-based encoder with 1.6B (1.6 billion) parameters. Mistral states that the model supports more than 160 languages, including all official EU languages.
The reasoning depth can be set to "low," "medium," or "high." The default is "high," and reasoning cannot be turned off entirely.
According to Hugging Face, the license is a modified version of the MIT License. Per VKTR's reporting, providers who offer the model as a service to third parties are required to obtain a commercial license from Mistral.
The previous model, Large 3, was licensed under Apache 2.0. Compared to that, service providers face slightly more restrictions. If you are only using the model for your own internal operations, the impact appears to be limited.
According to VKTR, API pricing is $0.70 per million input tokens and $2.10 per million output tokens — roughly half the preview pricing of $1.36 input / $4.18 output.
According to Hugging Face's announcement, the model's strongest areas are cybersecurity, coding agents, and multimodal capabilities (handling both images and text).
A word of caution here: the majority of these figures were produced by Mistral and its partners.
Indeed, Pasquale Pillitteri's article reports a score of 48.05% ranking 32nd out of 44 models on the third-party Vals v2.1 benchmark. On Terminal-Bench 4, Mistral's published figure is 28.3%, while the independent evaluation came in at 22.73%.
If the announced numbers hold up, this is a powerful model. Even so, it is safer to test it on your own business data before drawing conclusions.
Based on VKTR's summary of Mistral's own comparisons, the positioning looks like this:
Mistral claims to be "the best open-weight model from the West" on the Artificial Analysis index. However, it has also been reported that leading Chinese models score higher overall.
In short, rather than claiming across-the-board superiority, Mistral's positioning is as a Western-origin open model with particular strength in cybersecurity.
No Japanese-language benchmarks have been published at this time. Despite support for 160+ languages, you will need to test Japanese-language performance yourself.
Consider, for example, an IT department at a manufacturer that cannot send confidential data outside the company. The ability to host the model on an internal server is a major advantage of open-weight models.
However, even a quantized version requires 512 GB of memory or multiple GPUs. Some reports estimate 4 to 8 NVIDIA B200 or B300 GPUs. Running this independently is not realistic for individuals or small companies.
In that case, the options are to use the API or rent a cloud-based inference service. Hugging Face notes a configuration of 8 NVIDIA H200 GPUs achieving approximately 65 tokens per second per request.
For security professionals, the ability to analyze code vulnerabilities in-house is appealing. Mistral emphasizes that the model is well-suited for testing that large commercial AI providers tend to decline.
The weights are free to download. However, you will need to pay for the computing infrastructure to run them. A commercial license is also required if you plan to offer the model as a service to others.
It is difficult. Even a quantized version requires 512 GB of memory or multiple GPUs. The official NVFP4 version is designed for NVIDIA Blackwell-generation GPUs.
It is said to support 160+ languages, but no official figures for Japanese-language performance have been published. Please test it with your own documents before using it for business purposes.
Many of the figures were published by Mistral or its partners. Independent evaluations have produced lower results in some cases. Now that the weights are public, third-party verification is expected to progress going forward.
No. You can choose from three levels — low, medium, or high. The default setting is "high."
Start small — try the API with your own Japanese-language data first.
This article is a cross-post from AI Friends.
