IT之家 10 月 6 日消息,当地时间 10 月 6 日,法国公司 Mistral AI 正式推出 Mistral Large 4 模型的公开预览版(Public Preview),简称 ML4,代号 le Chonk。
即日起,使用者可以在 Mistral Studio 上感受预览版 API,模型权重将于本月底正式开放下载。
当下,官方正联合顶尖网络安全机构、受信任的合作伙伴及政府监管部门,在真实业务环境中对模型展开红队测试(Red-Teaming)。相关合作方将取得该模型的专项测试访问权限;该测试版本适当放宽了内容安全审查策略,并强化了网络安全攻防工艺能力。
ML4 是一款具备 1 万亿总参数、490 亿激活参数的原生多模态模型,Mistral 称实际上测表现不仅足以媲美全球顶尖的开源模型,更大幅领先于欧美开发的所有开放权重模型。
在网络安全、金融与法律等关键企业级工作负载中,该模型在所有开源模型中均达到了 SOTA。
在某些特定领域(例如视觉定位),它的表现更进一步,甚至超越了当前的前沿闭源模型。
据介绍,ML4 是在 Mistral 位于欧洲的自建数据中心内,基于 3,800 块英伟达 Grace Blackwell GPU 从零开始训练完成的,其欧洲部署思路完全由 Mistral 实施端到端独立运营,严格遵循欧盟法律法规,不受制于任何其他第三方数字服务给出商。
ML4 的训练数据具备显著的多语言特性,语料覆盖超 160 种自然语言,全面囊括了欧盟的所有官方语言。
售价维度,ML4 每百万 tokens 输入 1.36 美元(IT之家注:现汇率约合 9.1 元人民币),每百万 tokens 输出 4.18 美元(现汇率约合 28.1 元人民币)。
IT之家附相关分数如下:
本文来源:https://www.ithome.com/1/010/108.htm
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。
Thanks to IT Home reader fuhao for the tip!
IT Home October 6 news — On October 6 local time, French company Mistral AI officially launched the public preview (Public Preview) of the Mistral Large 4 model, abbreviated as ML4, codenamed le Chonk.
Starting today, users can experience the preview API on Mistral Studio, and the model weights will be officially available for download at the end of this month.
Currently, the official team is working with top cybersecurity agencies, trusted partners, and government regulators to conduct red-teaming tests on the model in real-world business environments. Related partners will obtain special testing access to the model; the test version appropriately relaxes content safety review policies and strengthens cybersecurity offensive and defensive process capabilities.

ML4 is a native multimodal model with 1 trillion total parameters and 49 billion activated parameters. Mistral claims that its actual performance not only rivals the world's top open-source models but also significantly surpasses all open-weight models developed in Europe and the United States.
In key enterprise workloads such as cybersecurity, finance, and law, the model achieves SOTA across all open-source models.
In certain specific areas (e.g., visual localization), its performance goes even further, surpassing current leading closed-source models.
According to introductions, ML4 was trained from scratch within Mistral's self-built data centers in Europe, using 3,800 NVIDIA Grace Blackwell GPUs. Its European deployment is independently managed by Mistral from end to end, strictly complying with EU laws and regulations, and is not subject to any other third-party digital service providers.
ML4's training data features significant multilingual characteristics, with corpora covering over 160 natural languages, comprehensively including all official languages of the European Union.
In terms of pricing, ML4 costs $1.36 per million input tokens (note: approximately 9.1 RMB at current exchange rates) and $4.18 per million output tokens (approximately 28.1 RMB at current exchange rates).

IT Home provides the relevant scores below:







Source: https://www.ithome.com/1/010/108.htm If there are copyright issues, please inform us by email: 13951431913#139.com or leave a comment on the right.