Fastino releases agent-tuned Nemotron finance, healthcare models
Fastino Labs has released two open-weight finance and healthcare models built on NVIDIA's Nemotron 3.5 Lightning, and it says an autonomous agent fine-tuned both with no human-run experiments.
Fastino Labs announced Fastino-Nemotron-3.5-Lightning-Finance and Fastino-Nemotron-3.5-Lightning-Healthcare, which it built with NVIDIA's Nemotron team. Both are on Hugging Face under the Apache 2.0 license and launched alongside NVIDIA Nemotron 3.5 Lightning, a 30B-parameter mixture-of-experts model with 3B active parameters.
Fastino reports that the finance model raises FinQA execution accuracy from 15.86% for the base model to 59.23%. Its other reported gains include:
- TAT-QA F1: 19.01% to 56.63%
- FinEntity macro-F1: 60.16% to 79.54%
- SEC-Num: 79.74% to 87.60%
The company says the healthcare model shows "verified gains" on MEDEC and MedCalc-Bench and improves on eight medical benchmarks in total. The excerpt available does not include its specific healthcare scores. Fastino also says both models improved on related tasks they were not trained for, and it presents this as evidence of real domain capability rather than benchmark fitting. None of these results have been independently checked.
Fastino says both models came from its Fastino Fine-Tuning Agent, which is launching in private preview. According to the company, the agent took a natural-language prompt and handled the rest: it analysed the base model's failure patterns, sourced and curated training data, ran training configurations in parallel, and ran its own contamination and transfer tests. Fastino says the process took less than 10 hours with no human in the loop. NVIDIA helped write model-specific fine-tuning recipes that gave the agent a head start.
The company is pitching the approach at regulated industries, where proprietary data has to stay inside the organisation. It says the agent can also fine-tune other open models on a customer's own data.