Fine-tuned 30B NVIDIA Model Beats 50x Larger Models in Code Verification

fhuszar · x · 2026-08-12

The team at Reasonable fine-tuned NVIDIA's latest open-weight 30B model, Nemotron 3.5 Lightning, to write machine-checked software correctness proofs in Verus.

After training on 4B tokens of synthetic data, the model outperforms a model 50x its size on per-attempt pass rate and nearly matches it on pass@3, while generating tokens faster than other similarly sized open models. The writeup also explores how models fail at formal proofs, how they attempt to cheat, and why high throughput is critical for interactive verification and CI pipelines.

Original post →

More from coding & agent

coding & agent channel →