Mistral发布万亿参数模型权重，主打网络安全与本地部署

法国人工智能公司Mistral发布了其万亿参数开源模型Mistral Large 4（简称ML4，昵称“Le Chonk”）的权重 1。该模型可在Hugging Face和API上获取，允许企业和政府自行下载部署 1。此举旨在通过提供可本地部署的前沿AI能力，增强网络安全防御主权 1。

在性能对比方面，Mistral声称ML4相比中国开源竞品GLM-5.3具有优势，每个任务少用47%的token，成本更低 1。此外，Mistral称其在Artificial Analysis的Coding Agent Index基准测试中，代理编码任务表现优于所有其他开源模型 1。Mistral科学副总裁Pierre Stock表示，开源权重确保访问权限无法被切断，有助于实时识别和响应重大安全事件 1。

Mistral CEO Arthur Mensch提出，AI应作为“赋能而非主导的工具”（a tool for empowerment, not dominance） 1。这一举措强调了对数据和系统控制权的重视，并加剧了开源模型与闭源模型提供商之间的竞争 1。

French artificial intelligence company Mistral has released the weights for its trillion-parameter open-source model, Mistral Large 4 (ML4), also nicknamed "Le Chonk," enabling enterprises and government bodies to download and deploy the system locally 1. The model is accessible via Hugging Face and through API interfaces, supporting on-premises installations that allow organizations to maintain direct control over their data and systems 1. By offering these frontier AI capabilities for local deployment, Mistral aims to enhance cybersecurity defense sovereignty, ensuring that access to the model cannot be arbitrarily revoked 1.

In terms of performance and efficiency, Mistral claims that ML4 consumes 47% fewer tokens per task compared to its Chinese open-source competitor, GLM-5.3, resulting in lower operational costs 1. Additionally, the company states that ML4 outperforms all other open-source models in agentic coding tasks according to the Coding Agent Index benchmark by Artificial Analysis 1. Pierre Stock, Mistral’s Vice President of Science, emphasized that the availability of open weights facilitates real-time identification and response to major security incidents 1. Arthur Mensch, CEO of Mistral, described the initiative as positioning AI as "a tool for empowerment, not dominance" 1. This strategic move intensifies the competition between open-source providers and closed-model vendors, underscoring a growing industry focus on data and system control 1.