Six months after its last nine-figure funding round, Harvey AI Corp. today announced that it has raised $550 million at $15.5 billion valuation.
Diffusion and Lightspeed Venture Partners led the deal. They were joined by more than a dozen other backers, including Sequoia, Kleiner Perkins and Goldman Sachs.
Harvey sells a cloud platform that law firms use to automate manual work for their attorneys. The company says its installed base includes 80% of the 100 highest-ranked law firms in the U.S. Harvey is also popular among large enterprises’ in-house legal teams: It counts half the Fortune 10 as customers.
The company’s platform enables users to store up to 100,000 documents in a repository called Vault. A built-in search engine uses artificial intelligence to surface useful patterns. For example, an attorney could ask Harvey to find supplier contracts that must be modified to comply with a new regulation.
The process of drafting a new legal document often requires lawyers to review their company’s existing agreements. For example, an attorney might scan historical customer contracts to find clauses that should be included in an important new deal. Harvey’s AI helps users find relevant documents in Vault. It also retrieves external data such as precedents and legislative clauses.
Earlier this year, Harvey rolled out AI agents that enable users to automate more complex tasks. An investment firm, for example, could use the agents to find potential issues in a large collection of due diligence documents. The agents ask attorneys for input when they need clarification about a tricky part of a task.
Today’s funding milestone comes a few days after the company debuted its first custom large language model. Tenet is a fine-tuned version of Kimi K3, an open-source LLM with 2.8 trillion parameters. It comprises 896 individual neural networks that are each optimized for a different set of tasks.
Harvey fine-tuned Tenet by training it on legal documents. It also added a custom harness, a collection of prompts and other technical assets designed to optimize LLM output quality. Harvey says Tenet performs some contract processing tasks 20% better than Kimi K3. Furthermore, it outperforms Fable 5 and GPT-5 Sol in multiple areas.
Harvey stated in Tenet’s launch blog post that it plans to deploy more computing infrastructure to advance its AI research. In particular, the company will prioritize the development of “new generalist models.” Today’s funding round should make it easier for Harvey to shoulder the steep costs associated with custom model development.
Training custom LLMs could not only help the company differentiate its feature set but also boost its margins. Developing a proprietary model is expensive, but it can lower infrastructure costs in the long term by reducing the need to use external models. The reason is that inference usually accounts for a much bigger percentage of an AI workload’s cost than training.
Harvey debuted Tenet alongside LAB, a benchmark that measures LLMs’ ability to complete legal work. The current iteration of the evaluation includes about 1,200 tasks. Harvey plans to expand LAB to more jurisdictions and areas of law.
