Harvey Secures $550M at $15.5B as Its Legal AI Model Tenet Beats Kimi K3 by 20%

What Harvey Actually Sells

Harvey’s platform targets law firms and in-house legal teams with a focused set of capabilities: document storage, AI-powered search, and increasingly, autonomous agents.

Its Vault repository holds up to 100,000 documents. The built-in search engine surfaces patterns across that corpus—so an attorney can, for example, query which supplier contracts need updating to comply with a new regulation, rather than reviewing each one manually.

The company reports that 80% of the top 100 U.S. law firms are customers, along with half of the Fortune 10. That kind of penetration at the top of the market is a meaningful signal of enterprise trust, not just early adoption.

Agents Add a Layer of Automation

Earlier this year, Harvey introduced AI agents capable of handling more complex, multi-step tasks. An investment firm conducting due diligence, for instance, could deploy agents to flag potential issues across a large document set. When the task requires judgment, the agents pause and ask the attorney for clarification.

This human-in-the-loop design is deliberate. Legal work carries high stakes, and the agents are positioned as assistants that escalate rather than decide autonomously.

Tenet: Harvey’s First Custom LLM

A few days before the funding announcement, Harvey launched Tenet—its first proprietary large language model. Tenet is a fine-tuned version of Kimi K3, an open-source model with 2.8 trillion parameters. It is structured as 896 individual neural networks, each optimized for a distinct set of tasks.

Harvey trained Tenet on legal documents and added a custom harness—a curated set of prompts and technical assets designed to improve output quality in legal contexts. The result, according to Harvey, is a model that outperforms the base Kimi K3 by 20% on certain contract processing tasks. It also reportedly surpasses Fable 5 and GPT-5 Sol in multiple evaluation areas.

These are Harvey’s own claims, and independent verification matters here. That said, the performance gap on contract tasks—if it holds—is the kind of domain-specific advantage that justifies the investment in custom model development.

LAB: A Benchmark Built for Legal Work

Alongside Tenet, Harvey introduced LAB, a benchmark designed to measure how well LLMs perform on legal tasks. The current version covers approximately 1,200 tasks. Harvey plans to expand it to additional jurisdictions and areas of law.

Building the benchmark alongside the model is a strategic move. It lets Harvey define the evaluation criteria for legal AI performance—which shapes how the broader market assesses competing tools.

The Infrastructure Logic Behind the Round

Custom model development is expensive upfront, but it carries a long-term cost argument. Inference—running a model at scale—typically accounts for a far larger share of AI workload costs than training. A proprietary model reduces dependence on external providers, which can meaningfully improve margins over time.

Harvey has stated it will use the new capital to expand computing infrastructure and develop what it calls “new generalist models.” The funding makes that roadmap financially viable at the scale legal AI now demands.

What This Means for Legal Teams and AI Buyers

For law firms and enterprise legal teams already evaluating AI tools, Harvey’s trajectory raises a few practical questions:

• Vendor depth vs. flexibility: Harvey is building a vertically integrated stack—custom models, proprietary benchmarks, domain-specific agents. That depth can mean better performance, but it also means deeper lock-in.
• Benchmark credibility: LAB is Harvey’s own benchmark. Useful as a reference point, but worth comparing against independent evaluations as they emerge.
• Cost trajectory: As Harvey internalizes more of its model infrastructure, pricing dynamics may shift. Enterprise buyers should track whether that translates into more predictable costs or new pricing structures.

Harvey is no longer just a legal AI application layer. It is building toward being a full-stack legal AI infrastructure company—with the valuation, customer base, and model capabilities to back that positioning.