[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-harvey-tenet-kimi-k3-legal-model":3},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":24,"news_slug":31,"published_at":32,"created_at":33,"modified_at":34,"is_published":35,"publish_type":36,"image_url":14,"view_count":37},"d1e8997e-bb60-453d-9ef8-71b8bdde5386","Harvey 首个自研法律模型 Tenet 曝光:底座没选 GPT 和 Claude,选了 Kimi K3","据 Business Insider 独家报道,法律 AI 独角兽 Harvey 发布首个自研专有模型 Tenet:雇佣律师批量构造模拟纠纷与案卷作为训练数据,在月之暗面开源的 Kimi K3 上训练而成。动因是降低对 OpenAI\u002FAnthropic 的调用成本,并应对模型供应商亲自下场做法律产品的竞争。同期 Artificial Analysis 的 Harvey LAB-AA 法律基准上,公开版 Kimi K3 以 94.6% 的准则通过率排名第一。","# 从\"用别人的模型\"到\"训自己的模型\"\n\n一家把法律 AI 生意做到 110 亿美元规模的美国独角兽,要出自己的第一个模型了。而它的训练底座,既不是 GPT,也不是 Claude,而是中国公司月之暗面(Moonshot AI)开源的 Kimi K3。\n\n据 [Business Insider 独家报道](https:\u002F\u002Fwww.businessinsider.com\u002Fharvey-builds-tenet-ai-model-for-legal-work-2026-8),法律 AI 公司 Harvey 在周二发布了其首个自研专有模型 Harvey Tenet,定位是承担以往需要律师花数小时甚至数天才能完成的工作,并且成本低于它目前依赖的第三方模型。\n\n## Harvey 是谁,为什么突然要自研\n\nHarvey 的原有生意模式,是典型的\"模型层之上的应用层\":在 OpenAI、Anthropic 等公司的通用模型之上,搭一套面向律所和企业法务的软件。Business Insider 称其已经做成了一门 110 亿美元规模的法律软件生意。\n\n问题在于,每次律师通过 Harvey 调用一次外部模型,Harvey 都要向模型供应商付费,用量越大这笔\"过路费\"越重。更微妙的是,Anthropic 正在用文档审查、起草类插件直接争取律师客户,OpenAI 则雇了 Ironclad 创始人 Jason Boehmig 领军法律业务——**供应商正在变成竞争对手**。\n\n自研模型的动机因此非常直白:把更多任务路由到自己的引擎上,在不向客户涨价的前提下改善毛利。联合创始人 Gabe Pereyra(前 Google DeepMind 研究员)同时强调质量动因:Harvey 本来就在按任务把工作分发给最擅长的模型,Tenet 是这个路由组合里的新选项。\n\n## 训练数据是\"雇律师造出来的\"\n\n要训一个法律模型,先得有能教模型\"像律师一样思考\"的数据。BI 披露的做法是:Harvey 雇佣了在职和签约律师(通过 Mercor、Snorkel 这类公司签约),让他们**虚构模拟纠纷和案卷**,再给模型在这些材料上的推理表现打分,用这套流程造出的数据去训练模型。\n\n而训练的底座,BI 的表述是:Harvey 用这些材料训练了 Kimi K3 的一个版本——\"一个来自中国初创公司月之暗面的低成本开源模型,自 7 月发布以来,该模型凭借能力和价格在科技界掀起了热潮\"。\n\n[MoonshotAI 的 GitHub 仓库](https:\u002F\u002Fgithub.com\u002FMoonshotAI\u002FKimi-K3)显示,Kimi K3 是一个 2.8 万亿参数的开放权重模型,基于 Kimi Delta Attention(KDA)与 Attention Residuals 架构,原生多模态,支持 100 万 token 上下文窗口。\n\n## 公开版 Kimi K3 本来就是法律基准第一\n\nHarvey 选 Kimi K3 并不突然。在 Artificial Analysis 独立运行的 [Harvey LAB-AA 法律基准](https:\u002F\u002Fartificialanalysis.ai\u002Fevaluations\u002Fharvey-lab-aa)上——120 个私有任务、覆盖 24 个法律执业领域,agent 需要在沙盒里通读案卷并产出备忘录、披露清单、质证摘要等真实法律交付物——公开版 Kimi K3 (max) 以 94.6% 的准则通过率(criteria pass rate)排名第一,领先 Claude Fable 5(93.6%)和 Muse Spark 1.1(93.1%)。\n\n换句话说,在 Harvey 自己参与出的考卷上,最强考生之一本来就已经是 Kimi K3。\n\n但有两个保留意见:其一,**Tenet 本身的基准分数尚未发布**,Harvey 只说\"很快会放出对比研究\";其二,BI 自己也提醒,模型厂商的自测基准天然要打折扣——厂商先用测试找出短板再针对性训练,时间长了有点像\"帮着写完答案再考试\"。\n\n## 不止一个模型,是一整套 Harvey II\n\nTenet 只是被包装进\"Harvey II\"的一部分。首席产品官 Anique Drumright 介绍,Harvey 还上线了新的 Memory 功能,让用户保存自己的工作偏好,agent 在跨任务执行时可以携带这些指令。\n\nPereyra 的终局设想更激进:让 Tenet 成为律所训练**自有模型**的起点底座——每家律所在 Tenet 之上,用自己律师数十年的工作方法继续训练出专属版本。BI 的评论是,这会让 Harvey 越来越像\"四大\"那样的专业服务公司而非软件供应商;而一家长期被讥讽为\"ChatGPT 套壳\"的公司,如果真做成,**套壳反而可能变成整个技术栈里最值钱的那一层**。\n\n值得注意的时间细节:Tenet 目前还没有上线到 Harvey 产品里,公司也拒绝透露具体时间,Pereyra同样拒绝点名哪些律所可能在内测。\n\n## 所以呢\n\n对中国开源模型来说,这件事的信号比任何榜单都硬:下载量只能说明模型\"被拿走\",而被一家美国独角兽选为自研模型的训练地基,才说明它\"被依赖\"。当垂直玩家连自研模型的地基都敢交给中国开源模型时,开源模型的全球化叙事,才算真正从 Hugging Face 的榜单,走进了别人的生产管线和成本结构。","https:\u002F\u002Fwww.businessinsider.com\u002Fharvey-builds-tenet-ai-model-for-legal-work-2026-8","6e1b5ecb-cb95-4c11-9d4e-6e6cd8d11a70",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"6ad31a14-c0da-42df-81fd-564281f768db","agentic-ai",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"a8002d98-9df1-4ab9-94d4-a7625af634c4","china-ai",{"id":19,"name":20,"slug":20,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"id":22,"name":23,"slug":23,"description":14,"color":14},"7e89b5cc-57db-4f37-bc6d-28919a73931c","model-release",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"23057cee-7531-4fc6-8b0c-2739ee256661","en","Harvey's Tenet builds on Kimi K3, not GPT or Claude","Legal-AI unicorn Harvey has unveiled Tenet, its first in-house proprietary legal model — trained on a version of Kimi K3, the open-weight model from China's Moonshot AI, using mock disputes and case files created by hired attorneys. The move targets inference costs and suppliers-turned-competitors like OpenAI and Anthropic. On Artificial Analysis's Harvey LAB-AA legal benchmark, the public Kimi K3 already ranks first with a 94.6% criterion pass rate.","# From \"Using Other People's Models\" to \"Training Their Own\"\n\nAn American unicorn that built an $11 billion legal-AI business is about to ship its first model — and the training base is neither GPT nor Claude, but Kimi K3, the open-weight model released by Chinese startup Moonshot AI.\n\nAccording to a [Business Insider exclusive](https:\u002F\u002Fwww.businessinsider.com\u002Fharvey-builds-tenet-ai-model-for-legal-work-2026-8), legal-AI company Harvey on Tuesday introduced Harvey Tenet, its first in-house, proprietary model for legal work. It is designed to take on more of the tasks typically done by lawyers over hours or days, at a lower cost than the third-party models Harvey currently relies on.\n\n## Who Is Harvey, and Why Build a Model Now\n\nHarvey's original playbook was the classic \"application layer on top of the model layer\": software for law firms and corporate legal teams built on general-purpose models from OpenAI and Anthropic. Business Insider reports it has grown into an $11 billion legal-software business.\n\nThe problem: every time a lawyer calls an external model through Harvey, Harvey pays the model provider — and that toll compounds with usage. The more uncomfortable part is that suppliers are turning into competitors. Anthropic has been chasing lawyers with plugins for document review and drafting, while OpenAI hired Ironclad founder Jason Boehmig to lead its legal push.\n\nThe motivation for an in-house model is blunt: route more work through your own engine and improve margins without asking customers to pay more. Cofounder Gabe Pereyra, a former Google DeepMind researcher, also cited quality: Harvey already routes different tasks to different models based on their strengths, and Tenet becomes another option in that mix.\n\n## The Training Data Was \"Manufactured by Hiring Lawyers\"\n\nTo train a legal model, you first need data that teaches it how lawyers think. BI's reporting lays out the method: Harvey hired attorneys, on staff and on contract through companies like Mercor and Snorkel, to **dream up mock disputes and case files**, then graded the models on how well they reasoned through them — and trained on the result.\n\nAs for the base, BI's wording: Harvey used the material to train a version of Kimi K3 — \"a low-cost, open-source model from the Chinese startup Moonshot\" that has whipped the tech world into a frenzy over its power and price since its July release.\n\n[MoonshotAI's GitHub repository](https:\u002F\u002Fgithub.com\u002FMoonshotAI\u002FKimi-K3) describes Kimi K3 as a 2.8-trillion-parameter open-weight model built on Kimi Delta Attention (KDA) and Attention Residuals, natively multimodal, with a 1-million-token context window.\n\n## The Public Kimi K3 Was Already #1 on the Legal Benchmark\n\nHarvey's choice of Kimi K3 didn't come from nowhere. On Artificial Analysis's independent [Harvey LAB-AA benchmark](https:\u002F\u002Fartificialanalysis.ai\u002Fevaluations\u002Fharvey-lab-aa) — 120 private tasks spanning 24 legal practice areas, where agents read case files in a sandbox and produce real legal deliverables such as memos, disclosure schedules, and deposition summaries — the public Kimi K3 (max) ranks first with a 94.6% criterion pass rate, ahead of Claude Fable 5 (93.6%) and Muse Spark 1.1 (93.1%).\n\nIn other words: on the exam Harvey itself helped write, one of the strongest test-takers was already Kimi K3.\n\nTwo caveats, though. First, **Tenet's own benchmark scores have not been released** — Harvey only says research comparing it to other models is coming soon. Second, BI itself cautions that vendor-run benchmarks deserve skepticism: makers use tests to find weaknesses and then train against them, which over time is a bit like taking an exam after helping write the answer key.\n\n## Not Just a Model — a Full \"Harvey II\" Rollout\n\nTenet is part of a broader rollout the company calls Harvey II. Chief Product Officer Anique Drumright says Harvey is also adding a new \"Memory\" feature that lets users save preferences about how they work, so agents can carry those instructions across tasks.\n\nPereyra's endgame is more aggressive: he wants Tenet to become the starting point for law firms to train **their own models** — each firm layering decades of its lawyers' know-how on top of Tenet to build a proprietary version. BI's take: that would push Harvey toward looking less like a software provider and more like a Big Four professional-services firm. And for a company long dismissed as a \"ChatGPT wrapper,\" if it pulls this off, **the wrapper starts to look like the most valuable layer in the stack**.\n\nOne timing detail worth noting: Tenet isn't live in Harvey's product yet, the company won't say when it will be, and Pereyra declined to name any law firms that might be testing it.\n\n## So What\n\nFor Chinese open-weight models, this signal is harder than any leaderboard. Download counts only prove a model was *taken*; being chosen as the training foundation of an American unicorn's first proprietary model proves it is *depended upon*. When vertical players are willing to build even their in-house model on a Chinese open-weight base, the globalization story of open models has genuinely moved from Hugging Face leaderboards into someone else's production pipeline and cost structure.","harvey-tenet-kimi-k3-legal-model","2026-08-18T17:30:00Z","2026-08-18T17:06:53.638852Z","2026-08-18T17:06:53.638860Z",true,"agent",207]