ZDNET的关键外卖
- 按计划，Mistral拥有其ML4型号的重量。
- ML4为西方开放型号设置了新栏。
——开放模式与专有模式竞争日益激烈。
法国AI实验室Mistral运出其最新型号"Mistral"大4 (ML4)，外号"Le Chonk"，为"万亿参数"，于10月6日预览。 实验室拥有发布模型的重量
开放权重 vs. closed: AI内战的脚步
新编码性能排名
ML4在代理编码基准测试 Terminal-Bench 4.0的得分从预览版的28%跳到41%，比Opencode的GLM-5.3低了一分，但很重要。 然而，在总指数上，一个模型的显示在三个相关的基准测试的综合分数上，ML4是GLM-5.3后的一个分数，得分48到GLM-5.3的49分。 编码GPT-6.1 Terra也总得分48分。
作为提醒，基准测试的分数本身应该用一粒盐来取出，特别是考虑到有多少模型作弊。
然而，这种业绩的改善对仅仅几个星期的额外培训是有意义的。 因为专有AI仍然昂贵，ML4“在人工分析编码Agent指数上采用GLM-5.3等顶级开放模型，同时平均使用47%的符号，”Mistral 称。
网络安全力量
当本月早些时候预览中的Mistral 发布 ML4时，它讲述了该模型如何优先考虑网络防御能力，广告可自定义的控制和数据主权。 该公司在过去几周里与开发商和领导者合作评估模型的网络安全能力，美国专有实验室如OpenAI、Google和Anthropic对于模型安全座标的担忧日益普遍。
类似Anthropic的玻璃翼项目和OpenAI推出的Astra,Mistral的初始测试伙伴也获得了一个保护性较弱的具有“扩大网络安全能力”的ML4版本。
至少在某种程度上，ML4的诺言似乎已经实现：Le Chonk在网络Gym 基准测试上为防御性网络安全评分刚刚高于DeepSeek V4.1 Flash，开源项目，复制其概念证明投入，并在不突破项目测试的情况下加以补齐。
最近AI安全事件使公开和专有模式相互对立。 最初由于其可商性被设定为不太安全，开放模型在Anthropic和OpenAI的专有模型被证明同样有风险后，被视为网络防御的新选择。
Mistral 称 ML4是为维持在用户控制下的安全而建造的——不同于专有模型，前沿实验室可以在任何时候技术上取消访问。
Mistral联合创始人Guillaume Lample 称在为10月6日预览发布作简报时
Hugging Face突破后，Mistral是许多签署Nvidia开放安全AI联盟的公司之一，这一跨行业伙伴关系主张开放模式，对于防御日益常见的AI安全事件的民主化至关重要。
还：这种新的ChatGPT骗局骗你安装恶意软件
通过Nvidia的逻辑及其旨在将AI安全工具民主化的联盟，比赛是在Mistral和其他开放模式之间，以实现最先进安全能道。
比较少计算
对像Mistral这样的三相实验室来说
还：这些 Claude 企业特性让你追踪AI ROI
Nvidia Grace Blackwell GPU在两个月内被从头开始训练，部署在欧洲的Mistral自己的数据中心。 对于上下文，NvidiaCEOJensen Huang 称在X上，OpenAI对GPT-6 Astra进行了约10万GPU的培训。
本文由 AI宝库自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
ZDNET’s key takeaways
- As planned, Mistral has released the weights for its ML4 model.
- ML4 sets a new bar for Western open models.
- Open models are increasingly competitive with proprietary models.
French AI lab Mistral shipped its latest model, Mistral Large 4 (ML4), nicknamed “Le Chonk” for its trillion-parameter size, in preview on Oct. 6. As promised, the lab has released the model’s weights, alongside some new benchmark scores that show Western open models gaining ground over their Chinese peers.
Also: Open weights vs. closed: An AI civil war’s afoot
New coding performance ranking
ML4’s score on the agentic coding benchmark Terminal-Bench 4.0 jumped to 41% from its preview version’s 28% to 41%, beating Opencode’s GLM-5.3 by one small but important point. On the overall Index, however, which shows a model’s composite score across three relevant benchmarks, ML4 is a point behind GLM-5.3, scoring a 48 to GLM-5.3’s 49. Codex GPT-6.1 Terra also scores a 48 overall.
As a reminder, benchmark scores themselves should be taken with a grain of salt, especially considering how many models cheat.
Still, that improvement in performance is meaningful for just a few weeks of extra training. As proprietary AI remains expensive to use, economical models that don’t sacrifice performance are eye-catching. ML4 “rivals top open models such as GLM-5.3 on the Artificial Analysis Coding Agent Index while using 47% less tokens per task on average,” Mistral said.
Cybersecurity prowess
When Mistral released ML4 in preview earlier this month, it touted how the model prioritizes cyber defense capabilities, advertising customizable control and data sovereignty. The company has spent the last few weeks working with developers and leaders to assess the model’s cybersecurity capabilities, a practice that’s becoming commonplace for proprietary American labs like OpenAI, Google, and Anthropic as concerns about model security mounts.
Similarly to Anthropic’s Project Glasswing and OpenAI’s rollout of Astra, Mistral’s initial testing partners got access to a less guardrailed version of ML4 with “expanded cybersecurity capabilities.”
At least by some measure, ML4’s promises seem to have come true: Le Chonk scored just above DeepSeek V4.1 Flash on the CyberGym benchmark for defensive cybersecurity, which measures how well models find “real memory-safety vulnerabilities in C/C++ open-source projects, reproducing them with proof-of-concept inputs, and patching them without breaking the projects’ tests.”
Recent AI security incidents have somewhat pitted open and proprietary models against each other. Initially framed as less safe because of their malleability, open models are viewed as a new option for cyber defense after proprietary models from Anthropic and OpenAI proved just as risky.
Mistral said ML4 is built for security that stays under user control — unlike proprietary models, which frontier labs can technically rescind access to at any time.
“The cyber defense capabilities will enable enterprises and governments to defend themselves against threat actors that are jailbreaking closed models to perform cyberattacks,” Mistral co-founder Guillaume Lample said during a briefing for the Oct. 6 preview release.
Following the Hugging Face breach, Mistral was one of many companies that signed Nvidia’s Open Secure AI Alliance, a cross-industry partnership that argues open models are crucial to democratizing defenses against increasingly common AI security incidents.
Also: This new ChatGPT scam tricks you into installing malware
By Nvidia’s logic, and its Alliance that aims to democratize AI security tools, the race is between Mistral and other open models to achieve state-of-the-art security prowess.
More for less compute
Training a competitive model in a compute shortage is no small task for a trimmer lab like Mistral, which doesn’t have the same resources as a pre-IPO giant like Anthropic.
Also: These Claude Enterprise features let you track AI ROI
“ML4 was trained from scratch on 4,000 Nvidia Grace Blackwell GPUs over two months, deployed in Mistral’s own data centers in Europe,” the company said. For context, Nvidia CEO Jensen Huang said on X that OpenAI trained GPT-6 Astra on roughly 100,000 GPUs.
Auto-fetched and translated by AI Gem Box. All rights belong to the original author.