Harvey Develops Its Own AI Model for Legal Software
Harvey introduces its own AI model to reduce dependence on OpenAI.
1Harvey's New AI Model: Tenet
Harvey has introduced Tenet, its own AI model designed for the legal field. This model was developed using post-training on Kimi K3, an open-weight model from Chinese startup Moonshot. With this move, Harvey aims to reduce its reliance on companies like OpenAI, Anthropic, and Google. Despite OpenAI being one of Harvey's investors, this step marks a significant move towards greater independence.
2Moonshot's Kimi K3 Model as the Foundation
The foundation of Tenet is Kimi K3, which was released by Moonshot in July 2023. Harvey trained this model specifically for the legal field in collaboration with Fireworks AI. The post-training process was designed to enhance the model's legal reasoning capabilities, allowing Harvey to offer a more tailored and specialized service to its clients.
3OpenAI's Investment and the Quest for Independence
Despite having significant investors like OpenAI, Sequoia, and Andreessen Horowitz, Harvey took a major step towards independence by developing Tenet. With a valuation of $15.5 billion, the company aims to reduce its dependence on OpenAI by using a Chinese-based model. This move is seen as part of Harvey's commercial strategy.
4Commercial Logic: Cost Control
Previously, Harvey had to pay competitors for every model call made on behalf of its clients. With the development of Tenet, these variable costs have been turned into fixed ones. This will help Harvey achieve long-term cost control. Additionally, using its own model could improve operational efficiency.
5Specialized Training for the Legal Field
Harvey worked with lawyers and law firms to train Tenet specifically for the legal field. Firms like Mercor and Snorkel created mock cases and files to evaluate how well models reasoned through them. This process helped enhance Tenet's legal reasoning abilities.
6European Competitor: Legora
Harvey's European competitor, Legora, is chasing a $10 billion valuation and selling to similar firms. Legora's choice of base model is as much a competitive issue as a technical one. Harvey's decision to use Kimi K3 could provide a significant advantage in this competition.
7The European Union's AI Act
The European Union's AI Act poses a significant question mark for companies like Harvey. When a company modifies someone else's general-purpose model, it is considered the provider of the modified version only if the change is significant. This could create uncertainty for Harvey's European clients.
8Licensing and Revenue Threshold
Kimi K3 allows derivative models but requires a separate agreement with Moonshot for model-as-a-service operators with revenues exceeding $20 million in any 12-month period. Harvey operates at over $350 million annualized, which could create additional licensing challenges.