Mistral AI发布开源权重万亿参数多模态模型Mistral Large 4

法国人工智能公司Mistral AI正式发布了其1万亿参数原生多模态模型Mistral Large 4（代号“Le Chonk”）的开源权重 1。此次发布旨在推动“主权AI”发展，允许企业在本地基础设施上完全控制、定制和部署模型，同时通过微软Foundry等平台提供托管服务选项 1。用户可通过Hugging Face下载权重，或在Microsoft Foundry、NVIDIA NIM等平台上使用 1。

在性能表现方面，该模型在代码生成基准测试中领先其他开源模型，在Cognition's FrontierCode 2.0中得分为19.7% 1。在Artificial Analysis Coding Agent Index中，其得分达48%，相比GLM-5.3少用47%的token 1；在Terminal-Bench 4.0中得分提升至41% 1。此外，在防御性网络安全基准Artificial Analysis Cyber Index中，该模型得分为51%，恶意请求拒绝率为96.6% 1。

为平衡性能与成本，Mistral Large 4引入了三种推理模式 1。API定价为输入每百万token 0.70美元，输出每百万token 2.10美元 1。该模型采用基于MIT修改的Mistral Ecosystem Open-Weight许可证，若涉及模型即服务（MaaS），商用需单独授权 1。

Mistral AI has officially released the open weights for its new large language model, Mistral Large 4, codenamed "Le Chonk." The model features one trillion parameters and is designed as a native multimodal system 1. This release supports the development of "sovereign AI," allowing enterprises to fully control, customize, and deploy the model on their local infrastructure 1. Developers can access the model weights directly via platforms such as Hugging Face 1.

The company reports that Le Chonk demonstrates leading performance among open-source models in specific technical benchmarks. On Cognition's FrontierCode 2.0 benchmark, the model achieved a score of 19.7% 1. In the Artificial Analysis Coding Agent Index, it recorded a score of 48%, requiring 47% fewer tokens than the competing GLM-5.3 model 1. Additionally, performance on Terminal-Bench 4.0 improved to 41% 1.

Security and defensive capabilities are also highlighted in the model's evaluation metrics. Le Chonk scored 51% on the Artificial Analysis Cyber Index, with a refusal rate of 96.6% for malicious requests 1. To balance performance and operational costs, the model introduces three distinct reasoning modes 1. When accessed through hosted services, pricing is set at $0.70 per million input tokens and $2.10 per million output tokens 1.

The weights are available under the Mistral Ecosystem Open-Weight license, which is modified from MIT terms; however, commercial use involving Model-as-a-Service (MaaS) requires separate authorization 1. Users can download the weights directly or utilize the model through integrated hosting platforms including Microsoft Foundry and NVIDIA NIM 1.