Fastino fine-tunes Nemotron 3.5 Super VL with tiny adapter
Fastino says a 75M-parameter LoRA adapter improved NVIDIA's 120B Nemotron 3.5 Super VL on all 15 finance and healthcare benchmarks it measured, while sharply cutting output tokens and therefore inference cost.
Fastino reported results from fine-tuning an early-access checkpoint of NVIDIA's Nemotron 3.5 Super VL, the model we covered at release. Super VL is a 120B-parameter hybrid Mamba-Transformer mixture-of-experts model with 12B active parameters. All training ran through the Fastino Training API using LoRA adapters that train 0.06% of the model's parameters.
Two domain adapters, one for finance and one for healthcare, raised every held-out benchmark Fastino measured, including five they were never trained on. In finance, FinQA rose from 35.2% to 77.9% and TAT-QA from 68.5 to 87.8 F1. ConvFinQA, which contributed no training data, jumped from 22.3% to 76.3%. In healthcare, BC5CDR went from 63.8 to 84.8 and the untrained BioRED from 46.1 to 69.0. On several tasks, a single combined adapter beat the single-task specialists Fastino had built for Nemotron 3.5 Lightning.
The company says the main gain is efficiency. On PubMedQA, the tuned model answered in 113 tokens instead of 1,597 and was 8.4 points more accurate. The stock Tinker cookbook math RL recipe lifted MATH-500 from 84.5% to 91.4% in 40 steps and made answers 25% shorter.
Costs were small. With 500 FinQA rows, 15 minutes on one B200 and about $2 of training, FinQA went from 35% to 77%. A full domain adapter took 2 to 5.5 hours and cost $21 to $43 at list price. Fastino also flagged a tradeoff: single-task FinQA adapters reduced TAT-QA scores. The mixed finance adapter avoided that drop.
These figures come from Fastino and were measured on a pre-release checkpoint.