Harvey built its own legal AI model instead of renting one
Harvey's Tenet model is a bet that applied AI moats will sit in domain-specific systems rather than generic frontier intelligence.
Pillar guide
Human-Centered AIAI strategy and adoption grounded in human needs, organizational reality, and responsible design.
BREAKING: Harvey built its own legal AI model instead of renting one from a frontier lab.
Tenet is Harvey’s first post-trained model, built on Moonshot’s open-weight Kimi K3 base and trained with Fireworks AI specifically for legal work. Harvey reports it hitting frontier-level results on its own Legal Agent Bench, at what the company describes as a fraction of the cost of the big labs, though it’s worth noting most of those benchmark numbers are self-reported rather than independently verified on a public leaderboard.
It’s not just a model, it’s a system built around three things: turning an 80-million-token M&A data room into a memo, extracting structured data across ten thousand-plus documents, and internalizing a firm’s own institutional knowledge so the model actually remembers how that specific firm works.
The cool signal for us is the strategy behind it: Harvey wants law firms to own specialized models rather than rent access to someone else’s. That’s a real bet that the moat in applied AI sits in the domain layer, not in whoever has the biggest general-purpose model. If it plays out, the companies most exposed are the legal AI wrappers whose entire pitch was “GPT plus a UI for lawyers.”
Worth asking yourself the same question about your own product: are you building on top of intelligence, or are you renting it with no plan for what happens when the rent goes up.