一家美国法律人工智能（AI）初创公司放弃美国模型，转用中国的Kimi K3。当地时间8月20日，总部位于旧金山的公司Harvey发布声明表示，该公司已基于Kimi K3模型构建了其首个内部模型Harvey Tenet。    While scientists and science fiction authors alike have long dreamed of sending crewed missions to Mars, a new study suggests these plans may hit a snag – a big one.,The authors of a yet to be peer-reviewed study argue that the participants of such a mission to Mars would end up being exposed to radiation – namely, to “energetic charged particles from Galactic and Solar origin” – both while in transit and while on the planet itself, with the “thin spacecraft shield” and “rarefied Martian atmosphere” offering little in the way of protection.,The research team attempted to gauge the impact of a simulated 1,000-day mission to Mars on various organs of the human body, and concluded that astronauts may be exposed to radiation doses above one Sievert, which is over the safe limit set by most space agencies, New Scientist points out.,“We suggest mitigation strategies such as improved ways of shielding and dietary supplements, and make recommendations for the safety of astronauts in future crewed missions to Mars,” the team wrote in the abstract of their work available on the arXiv preprint server.,Ideas about how astronauts may be protected from radiation during a mission to Mars have already been proposed, such as the concept of a radiation shield comprised of a fungus found growing at the Chernobyl Nuclear Power Plant.,Last year, one architecture studio even unveiled a project for a self-sufficient and sustainable city that would be built on the red planet itself.。香港《南华早报》21日报道称，此举凸显出，在开发成本飙升的背景下，西方科技公司正日益转向中国开源权重系统。此前，该公司主要使用Anthropic、OpenAI和谷歌等美国公司的封闭专有模型。Harvey公司在声明中表示，最新内部模型在复杂的法律工作中达到了“最先进”的性能。

B | 人工智能政策研究员西蒙·赫德林表示，Harvey的这一转向，是“开源模型的一个绝佳例子”。他在社交媒体发文指出，开源模型能使开发者能够在特定行业或企业数据上对系统进行后训练，从而提高准确性并降低推理成本。后训练（Post-Training）是指在预训练模型的基础上，针对特定的任务或数据集进行额外的训练。“我们仍处于探索高能力开源权重模型可能性的最早阶段，”赫德林说，“遗憾的是，美国在开发前沿开源权重模型方面落后了。”

贴文截图据介绍，Harvey公司成立于2022年，获得了OpenAI、红杉资本等美企支持，在今年3月的一轮融资中估值达到110亿美元，主要服务于大型国际律师事务所和企业客户。该公司在声明中提到，经过全面法律数据集训练的Harvey Tenet在一系列复杂、长周期的法律代理任务中，表现超过了其基础模型以及美国前沿系统——包括Fable 5和GPT-5.6 Sol。据Harvey称，这提高了成本效率。公司称，Harvey Tenet的训练使用了约150块英伟达B300图形处理单元（GPU），耗时两个月。报道提及，中国开源模型的高性价比，吸引了越来越多的西方企业。美国电信巨头ATamp;T副总裁马克·奥斯汀近期表示，目前尚未使用任何中国开源权重模型，但公司正在分析包括DeepSeek和月之暗面在内的中国公司的选项。