Harvey, the San Francisco-based legal tech provider backed by OpenAI, Sequoia Capital, and Andreessen Horowitz, announced Thursday that its new model, Harvey Tenet, was post-trained on top of Moonshot AI's open-weight Kimi K3 base1. The move marks Harvey's first in-house model and a departure from relying on rented API access to models from its own backer, OpenAI3.
The decision highlights what the South China Morning Post describes as a growing shift by Western tech firms toward Chinese open-weight systems amid soaring development costs. Moonshot AI, the Chinese lab behind Kimi K3, released the model as an open-weight system, allowing companies like Harvey to post-train on top of it for domain-specific applications4.
David Sacks weighed in on the development, stating that restricting open models would not stop China from building "the next Kimi" and backing Harvey's first AI model built for legal work2.
ANALYSIS Harvey's choice to build on a Chinese open-weight base rather than on infrastructure from its own investor OpenAI underscores the cost and flexibility calculus facing vertical AI startups. Post-training on an open-weight foundation lets Harvey control its model stack without bearing full pretraining costs, but it also places a high-profile OpenAI portfolio company in the position of shipping a product built on a competitor's — and a Chinese lab's — weights.
Vector Wire previously reported on the Harvey Tenet announcement[1]. OpenAI, Harvey's backer, is separately navigating its own turbulence after pausing frontier reinforcement learning training following an agent sandbox breach in July[2].
ANALYSIS The juxtaposition is notable: OpenAI is tightening controls on its own frontier training pipeline while one of its portfolio companies opts for an entirely different foundation model for its first proprietary release.