Harvey Launches Harvey II Platform and First Self-Trained Legal Model Tenet
On August 19, legal AI firm Harvey launched its second-generation platform Harvey II along with Harvey Tenet, the first proprietary model trained specifically for legal work. The platform organizes work around matters/projects, gives agents immediate access to files, context, permissions and history, adds memory and domain-specific context, and lets agents delegate tasks.
Confirmed
- Harvey II ships with matter-centric organization, memory features, and agent collaboration
- Per Business Insider (relayed by @meliarobin), Tenet is based on a version of Kimi K3, trained on simulated disputes and case files; co-founder Gabe Per was interviewed
- Training details from @meliarobin: post-training with FireworksAI using rank-64 all-network LoRA with importance-ratio masking, trained on 134 B300 GPUs for about two months
- Per @dejavucoder: training data mixes public legal data, synthetic data and expert data; Tenet's full-pass rate on the LAB benchmark rose 82% over the base model
Why it matters
- @meliarobin notes the in-house model helps Harvey shed its 'wrapper' reputation and cut costs — handling more work on owned infrastructure short-term, and serving long-term as a base module for law firms training their own models
- The move marks a flagship vertical player shifting from calling general models to training its own, and the disclosed training recipe offers a rare reference for vertical-model post-training
2026-08-19 ~ 2026-08-20 · 7 related posts
Primary sources
- [source] Legal AI firm Harvey launches first in-house model Tenet, built on Kimi K3 — meliarobin · 2026-08-19
- Harvey launches Harvey II with Tenet, its first model trained for legal work — saranormous · 2026-08-19
- Harvey launches Harvey II platform with memory and legal reasoning model — emmanuelvivier · 2026-08-20
- [source] Harvey's legal model Tenet: post-trained on Kimi K3 with 134 B300 GPUs for 2 months — meliarobin · 2026-08-20
- [source] Harvey releases Tenet legal model, boosts pass rates by 82% — dejavucoder · 2026-08-20
2 near-duplicate retellings: meliarobin · round