OpenAI-backed legal AI firm Harvey unveiled Tenet, its first in-house model, built on China’s open-weight Kimi K3.
Trained for two months on ~150 Nvidia B300s, Tenet lifted all-pass scores 82% over K3 on Harvey LAB, nearly doubled completed tasks and runs at under one-quarter of mainstream frontier-model inference cost.
Harvey processes 13T tokens monthly. Tenet won’t replace GPT-5.6 Sol or Claude Opus 5; it adds control and pricing leverage. The moat is shifting from base models to expert data and post-training.