Harvey 获得 5.5亿美元新融资，估值升至 155亿美元

Harvey 宣布 on 2026年9月9日 that it has raised $550 million at a $15.5 billion valuation in a funding round co-led by Diffusion and Lightspeed Venture Partners. The 法律 AI company said the capital will accelerate its efforts to help law firms, in-house legal teams, and professional services firms build and own their intelligence at scale.

Existing investors Sequoia, Kleiner Perkins, a16z, Coatue, Conviction, Elad Gil, Evantic, GIC, Goldman Sachs Alternatives, Verified Capital, and WNDR participated in the round. Diffusion, Lightspeed, Sapphire Ventures, and Whale Rock joined as new investors.

According to the announcement, the raise follows the introduction of Harvey Tenet, the company’s first 后训练的开放权重模型, and the launch of Harvey LAB（其法律代理基准）.

联合创始人声明与已报告的采用情况

Harvey 联合创始人 Winston Weinberg 和 Gabe Pereyra said the company “wants to be the global partner legal teams turn to for this work,” adding that Harvey plans to hire and develop its team to support clients through what they described as the next wave of AI transformation. The opportunity for companies to accelerate their competitive advantage with AI has never been higher, the co-founders said.

Harvey said 80% of Am Law 100 law firms use its products, alongside in-house legal teams that include five of the Fortune 10. Harvey’s 公司页面 describes the business as domain-specific AI for legal and professional services, with products spanning contract analysis, due diligence, compliance, and litigation workflows, and states that more than 2,400 customers in more than 70 countries use the platform. The page names GV, the OpenAI Startup Fund, Andreessen Horowitz, and EQT among the firm’s backers alongside Sequoia, Kleiner Perkins, and Coatue.

Harvey Tenet 研究预览

Harvey detailed Tenet in an 2026年8月20日 research preview. The model starts from a Kimi K3 base and was post-trained together with Fireworks research for long-horizon legal work, using asynchronous reinforcement learning with group-sequence policy optimization, or GSPO. The training dataset comprised roughly 1,750 agentic legal task environments, and training ran on approximately 150 NVIDIA B300 GPUs over two months.

Harvey reported that the model completes almost twice as many held-out LAB tasks as the base Kimi K3 and 20% more on LAB Contracts, increasing all-pass rates by 9 and 2 percentage points, respectively. According to the preview, Tenet achieves state-of-the-art performance on LAB Contracts and places second on LAB, and those gains generalized to external agent benchmarks the model had not seen during training, including Mercor’s APEX Agents corporate-law subset and Crosby’s Redline Bench.

The training corpus combined synthetic data, publicly available legal data, and human expert data, and Harvey said it worked with Mercor and others to build and scale the expert datasets. The company said it used reward shaping to incentivize efficient tool use and reasoning, preferring trajectories that reduce tokens consumed at inference time so that cost and performance could be co-optimized. Harvey stated that it did not use any customer data in its post-training efforts.

Beyond the core model, the preview describes separately trained capabilities Harvey said it will incorporate into its product over time: M&A diligence agents built on Recursive Language Model harnesses extended with Baseten research, Review Table extraction models trained with Applied Compute, and firm-knowledge models trained with Engram that encode a firm’s knowledge into parametric memory and structured notes.

开源法律代理基准

Harvey open-sourced the Legal Agent Benchmark on 2026年5月6日. Its first version contains more than 1,200 agent tasks across 24 legal practice areas, evaluated by more than 75,000 expert-written rubric criteria, and is available on GitHub. Each task pairs a short instruction written as a partner’s request for work with a client matter of documents and requires the agent to produce a reviewable work product, which is graded against expert rubrics broken into atomic pass/fail criteria covering facts, conclusions, citations, and formatting. A task is marked complete only when every criterion passes, a standard Harvey calls all-pass grading.

One benchmark task asks an agent to analyze change-of-control provisions for a fictional $458 million acquisition, review a virtual data room, assess deal risk, and prepare a draft memo for the deal team, with the work graded against 57 criteria spanning nine legal issues. Harvey said it released LAB to give model providers, agent builders, researchers, and law firms a shared way to measure progress on long-horizon legal agents. The benchmark launched without a leaderboard, and Harvey said at the time that it would work with research partners to publish baseline results and standards for normalized submissions.

In the August research preview, Harvey said its next steps include scaling LAB to more jurisdictions, practice areas, and workflows and expanding its compute to bring its research into production with new generalist models and capabilities.