Both versions are aligned block by block, in reading order: title, the essentials, then paragraph by paragraph. Where the translation merged or split a paragraph, the matching cell stays empty — we never pair two passages by guesswork.
美国法律人工智能公司Harvey的工程师把Kimi K3放进了一个新的法律模型Tenet里。Tenet面向的不是“法律知识问答”，而是律师每天要处理的长流程：从一堆客户文件中找线索，连续检索和调用工具，发现风险并补齐引用，最后交付合同修改稿或法律备忘录。\n\nHarvey成立四年，估值达到110亿美元，目前服务超过1300家机构和10万名律师。Tenet是它首个基于开放权重模型进行后训练的模型，专门应对需要长时间执行、连续调用工具的法律任务。公司在法律智能体基准LAB上的初步结果显示，Tenet成功完成的测试任务接近底座模型的两倍。\n\n这条路的关键，不是给Kimi换一个名字。通用模型也许知道法律知识，却未必知道一项真实工作应该从哪一步开始、什么时候需要再查一份材料，以及一份交付物究竟差在哪里。Harvey需要用自己的数据、算力和法律业务经验，把开放模型继续训练成更适合律师流程的工具。\n\n类似的“拿底座再加工”正在编程工具中出现。Cursor的Composer 2被确认基于Kimi K2.5训练，后续Composer 2.5仍沿用这一底座；英国Cosine用Kimi K2.6训练维护旧代码的Lumen Outpost，Devin的SWE-1.7也公开使用Kimi K2.7。短短半年，Kimi的不同版本依次进入这些海外产品。\n\n具体来说，这让中小型人工智能公司多了一条可行的商业路径：它们不用从零训练基础模型，也不用在产品越做越大时持续按调用量依赖OpenAI或Anthropic等闭源服务。客户得到的是法律、编程等更贴近工作流程的产品，而不是一个只会聊天的通用模型；模型提供方则可能成为其他产品的原料和上游底座。\n\n但目前能确认的仍是企业发布的结果。Tenet的“接近两倍”来自Harvey法律智能体基准LAB的初步测试；中国模型进入海外产品，也不等于这些产品可以省掉行业数据、算力和实际验证。开放权重降低了起点，真正的产品能力仍要在具体工作里训练出来。
Engineers at US legal AI company Harvey have put Kimi K3 into a new legal model called Tenet. Tenet is not designed for “legal knowledge Q&A,” but for the lengthy workflows lawyers handle every day: finding clues in piles of client documents, conducting repeated searches and calling tools, identifying risks, filling in citations, and ultimately delivering a revised contract or legal memorandum.
Harvey was founded four years ago, has reached a valuation of $11 billion, and currently serves more than 1,300 organizations and 100,000 lawyers. Tenet is its first model post-trained on an open-weight model, specifically to handle legal tasks that require extended execution and repeated tool calls. Initial results on the company’s legal agent benchmark, LAB, show that Tenet successfully completed nearly twice as many test tasks as the base model.
The key to this approach is not simply giving Kimi a new name. A general-purpose model may know legal concepts, but it may not know where a real task should begin, when another document needs to be checked, or what exactly is wrong with a deliverable. Harvey needs to use its own data, compute and legal expertise to continue training the open model into a tool better suited to lawyers’ workflows.
A similar form of “further processing of a base model” is emerging in programming tools. Cursor’s Composer 2 has been confirmed to be trained on Kimi K2.5, and the subsequent Composer 2.5 continues to use the same base. The UK’s Cosine trained Lumen Outpost, which maintains legacy code, on Kimi K2.6, while Devin has also publicly said it uses Kimi K2.7. In just six months, different versions of Kimi have successively entered these overseas products.
Specifically, this gives small and mid-sized AI companies another viable business path: they do not have to train a foundation model from scratch, nor do they have to keep relying on closed services such as OpenAI or Anthropic and paying by usage as their products grow. Customers get products for legal work, programming and other fields that are more closely aligned with real workflows, rather than a general-purpose model that only knows how to chat; model providers, meanwhile, may become the raw material and upstream foundation for other products.
What can currently be confirmed, however, is still limited to the results published by the companies. Tenet’s “nearly twice” figure comes from the initial testing of Harvey’s LAB legal agent benchmark; the entry of Chinese models into overseas products also does not mean those products can do without industry data, compute and real-world validation. Open weights lower the starting point, but genuine product capabilities still have to be trained through specific work.