Post-Trained Nemotron 3.5 Hits Opus-Level on Legal Tasks, Halves Token Use
sudoraohacker · x · 2026-08-12
Harvey, in collaboration with NVIDIA and Trajectory Labs, post-trained the Nemotron 3.5 Lightning model on the Legal Agent Bench (LAB).
Key findings from the post-training:
- Performance on held-out LAB tasks improved from 0% to 8.3%, achieving Opus-level all pass and even beating the much larger post-trained Nemotron 3 Ultra.
- Performance improved across nine practice areas with zero regressions.
- The model's average output length was reduced from 90k to 37k tokens, increasing the reward-per-token by 2.4x.
Related event: Fine-tuned NVIDIA Nemotron 3.5 Outperforms Claude Opus(4 posts)→
More from coding & agent
- Six-month prediction: hiring for generic agent harnesses built from scratch will cool — yi111 · 2026-10-16
- Omarchy ISO wouldn't boot on NVIDIA, so he had Claude patch it — indragie · 2026-10-16
- Dev uses Grok to clean up 3D printer firmware, catches it sneaking in nonsense code — yacineMTB · 2026-10-16
- A rude AI agent took 89% of a shared resource in 102-run multi-agent personality experiment — Brave-History-6502 · 2026-10-16
- Guardrails can make risky delegation work — but it takes a full-time team, says iandanforth — iandanforth · 2026-10-16
- What can you truly delegate? Only tasks with unambiguous, verifiable results — iandanforth · 2026-10-16