Fastino details its agent's Nemotron fine-tuning results
Fastino has published benchmark results for its agent-tuned Nemotron 3.5 Lightning finance model, showing a single rank-32 adapter more than quadrupling FinQA accuracy from about ten hours of autonomous search.
As we reported earlier, Fastino used its autonomous fine-tuning agent to turn early-access checkpoints of NVIDIA's Nemotron 3.5 Lightning into finance and healthcare models. A new Fastino research post by Dhruv Atreja now gives the numbers behind that release.
The agent treats fine-tuning as a search over three linked parts: the training data, the optimization settings (LoRA rank, learning rate, epochs) and the supervision strategy (for example, direct answers versus chain-of-thought, or which teacher model to distill from). It records each completed pipeline as a node in a graph and picks the next experiment with a Monte Carlo Graph Search–style policy. The method is described in an arXiv paper.
In finance, the agent fit six skills into one rank-32 adapter and kept about 87% of the gains of standalone specialists. FinQA execution accuracy went from 14.39% to 64.86%, and TAT-QA F1 from 17.19 to 74.87. Fastino says the base model largely failed these tasks because it could not produce the executable programs and structured answers they require. The adapter also improved on benchmarks it never trained on: ConvFinQA rose from 11.67% to 61.00%.
On the BigFinanceBench rubric, the model scored 40.28%. That is above Nemotron 3 Ultra (39.2%), Gemma 4 31B-IT (34.8%) and GPT-5.4 Mini (22.4%), and within 5 points of Kimi K2.6. Fastino notes that every open-weight model ranked higher activates at least 23B parameters per token. Its model activates 3B.
The available text cuts off before the healthcare results and the post's promised seven lessons and ten reusable recipes.