Harvey Trains Own Legal Model on Kimi K3, Signaling a Shift to Enterprise Self-Training

机器之心 · wechat · 2026-09-14

Jiqizhixin surveys the trend of application-layer AI companies training their own models, plus the launch of the PyTRIO training platform.

Industry cases

Legal AI company Harvey released its first open-weight post-trained model Tenet, built on Kimi K3, scoring 19.7% full pass rate on the legal agent benchmark LAB — above Fable5 (11.5%) and GPT-5.6 Sol (2.5%) — at a lower per-task cost of $5.92.
Cursor's Composer 2 (Kimi K2.5 base) uses continued pretraining and large-scale RL on real product feedback, hitting the top tier on coding benchmarks at 1/6–1/10 the inference cost of frontier coding models.
Bridgewater and Thinking Machines Lab fine-tuned Qwen3-235B with expert-labeled financial data, reaching 84.7% accuracy on six financial screening tasks vs 78.2% for the best frontier model, with 13.8× lower per-task inference cost.
Mercor CEO Brendan Foody frames this as enterprises gaining "autonomous intelligence": post-train open models per domain, then customize for key clients, reducing dependence on closed APIs.

PyTRIO platform

Emotional Machine's PyTRIO offers a Training-as-a-Service API: users control data, loss, reward and training loops locally; GPU compute, distributed training and environment setup run in the cloud. It has adapted Ascend 950PR/910B chips.
Example experiments: a $2 Search-R1 multi-turn search RL run lifted MacroEM from 28.57% to 45.71%; SAR-OPD medical distillation hit 84.33% on MedQA-zh without degrading general ability; VisionGRPO raised GeoQA accuracy from 71% to 87%. Token-based and GPU-hour billing fit different workloads.

More from Companies & People