Harvey's legal model Tenet: post-trained on Kimi K3 with 134 B300 GPUs for 2 months
meliarobin · x · 2026-08-20
Legal AI company Harvey shared an update on its post-training effort, a model called Tenet, built on Kimi K3 as the base in collaboration with FireworksAI.
Key details:
- Rank-64 LoRA over the full network, trained with GSPO using importance-ratio masking
- Compute: 134 B300 GPUs for 2 months
- Achieves SOTA on LAB and generalizes to third-party legal benchmarks (Mercor's Apex Agents Corporate Law, Crosby's Redline Bench, LegalBench) despite not being trained on third-party legal datasets
- Uses domain-specific sub-agents for complex tasks: M&A diligence via an RLM harness for long-horizon tasks (with Baseten), Review Table with specialist models for high-volume structured extraction (with Applied Compute), and firm knowledge capabilities
Related event: Harvey Launches Harvey II Platform and First Proprietary Legal Model Tenet(7 posts)→
More from Models
- Post-training causes LLMs to produce novel but impractical language — TuhinChakr · 2026-08-20
- User says Grok 4.6 now handles 100% of coding work previously done with Codex — CedricMakes · 2026-08-20
- Grok 4.6 leads in legal/GDP benchmarks, lags in coding — ChrisGPT · 2026-08-20
- Leaked System Prompt: Domestic Giant's Client Uses 3-Layer Memory & MCP Routing — vista8 · 2026-08-20
- Test finds base model completions inherit AI/human writing traits from prompts — AaronBergman18 · 2026-08-20
- Polymarket: 10% chance a Chinese model tops the leaderboard by year-end — Polymarket · 2026-08-20