Harvey’s First Proprietary Legal Model: Tenet
Introduction
Harvey has launched Tenet, its first proprietary model for legal work, post-trained on Kimi K3, an open-weight model from Chinese startup Moonshot. This development comes after years of using models from OpenAI, Anthropic, and Google.
The Model and Its Training
Tenet is based on the detail that matters: Kimi K3, an open-source model released in July by Moonshot, developed in collaboration with Fireworks AI. This choice is significant given OpenAI’s investment in Harvey, alongside Sequoia and Andreessen Horowitz.
Commercial Strategy
The commercial logic behind this move is straightforward. By owning its proprietary model, Harvey can reduce variable costs associated with model calls and turn them into fixed costs.
Quality Assurance
Gabe Pereyra, co-founder of Harvey and former Google DeepMind employee, emphasizes that Tenet offers a specific legal work option, tailored to suit various tasks. The training data was meticulously crafted by hiring lawyers from firms like Mercor and Snorkel to create mock disputes and case files, which were then evaluated for model performance.
European Competition
Legora, a Swedish alternative to Harvey, is also pursuing a significant valuation and targeting the same investors, making the choice of base model a competitive consideration as much as a technical one.
Legal Considerations
For European users, the AI Act poses questions regarding liability when modifying someone else’s general-purpose model. While a post-trained model may fall below the threshold for significant modification, most obligations still reside upstream. This means that deploying Tenet in Europe could involve a complex documentation chain starting in Beijing.
Open-Weight Debate and Privilege
The open-weight debate has shifted from capabilities and pricing to a question of who your counterparty is when the work is subject to attorney-client privilege.