Fastino unveils AI-built finance and healthcare models

Fastino developed Fastino-Nemotron-3.5-Lightning-Finance and Fastino-Nemotron-3.5-Lightning-Healthcare, two open weight models post trained on NVIDIA Nemotron 3.5 Lightning and available on Hugging Face. Both were produced by the Fastino Fine Tuning Agent, a fully autonomous autoresearch system that managed the entire pipeline in less than ten hours per model. No human at Fastino selected training data, tuned hyperparameters, or pushed the retry button after a failed run. The agent researched the task, curated data, built evaluations before training, ran multiple experiments in parallel, guarded against contamination, and delivered the strongest checkpoints. The company is opening the agent as a private preview for teams that want to build their own specialized models.

Why open weights matter in data sensitive domains

Finance and healthcare teams have been stuck between a rock and a closed model. The public FinQA leaderboard is topped by proprietary systems. HealthBench is dominated by models you cannot run locally. Meanwhile the data that would make a model good at the work that matters never gets to leave. Filings, patient notes, and internal metrics are governed for a reason. An open weight model changes that equilibrium. It lets you bring the model to the data, fine tune on premises or in your private cloud, and preserve the chain of custody and the IP. Nemotron 3.5 Lightning already carries a lot of domain knowledge. Targeted post training adds the precision you need for your specific workflows without handing the crown jewels to a third party.

Autonomous post training, not a research marathon

Fine tuning sounds simple until it becomes a tangle of choices you cannot separate. What data you train on, how you configure the run, and what supervision format you choose all interact. Grid searches do not capture those interactions, which is why manual post training often stretches into weeks. Fastinos agent treats the process as a guided search over whole pipelines. It reasons from one attempt to the next instead of wandering a parameter spreadsheet. Under the hood the search policy looks like a Monte Carlo Graph Search over recipes, data mixtures, and schedules. The agent also retains lessons from earlier jobs and can adapt to architectures that were off limits at first. That is why it could post train a newly released Nemotron model effectively.

What the numbers say

On finance, the gains are the sort you can feel in a real research stack. FinQA execution accuracy jumped from 15.86 percent to 59.23 percent, a lift large enough to cross the line from toy to tool. TAT QA climbed from 19.01 percent F1 to 56.63 percent. SEC Num rose from 79.74 to 87.60 percent. FinEntity improved from 60.16 to 79.54 percent macro F1. BizFinBench moved from 49.65 to 57.46 percent. Those are not curve fitted blips. The improvements transfer to related tasks the model was never explicitly trained for, which is what you want if you plan to trust it on the mess of real filings and the dense tables buried inside them.

Healthcare saw similar breadth. On HealthBench, which grades multi turn clinical conversations against physician written rubrics, the model stepped from 40.62 to 48.17 percent. On MedAgentBench overall success rate moved from 36.00 to 40.00 percent. Entity extraction improved markedly, with BC5CDR jumping from 47.92 to 72.51 percent and MedMentions from 19.74 to 40.29 percent. HealthAdminBench rose from 25.67 to 29.95 percent. On MEDEC, both flag accuracy and sentence accuracy climbed, from 53.66 to 64.98 percent and from 48.08 to 62.89 percent. On the full MedCalc Bench Verified set the model moved from 49.09 to 54.18 percent. HealthBench Pro and PubMedQA also saw steady lifts. As with finance, the gains carry into tasks the model was not trained for, which points to general domain capability rather than benchmark memorization.

Built on Nemotron openness

NVIDIAs Nemotron 3.5 Lightning is a 30 billion parameter mixture of experts model designed for long running agentic workloads. Crucially, Nemotron is open across weights, datasets, and training recipes. That openness gives an autoresearch agent something to explore. Every architectural decision that was published is a knob the agent can reason about. Every released dataset is a potential ingredient. Specialization at scale depends on that kind of transparency. A closed box cannot teach an agent what to try next.

Availability and practicalities

Fastino-Nemotron-3.5-Lightning-Finance focuses on numerical reasoning, research, and summarization across financial documents. Fastino-Nemotron-3.5-Lightning-Healthcare focuses on clinical conversations, documentation, and medical concept extraction. Both models are open weight and hosted on Hugging Face. They are released under the Apache 2.0 license. The Fastino Fine Tuning Agent is available in private preview, and teams can request access at Fastino dot ai.

Investors and context

Fastino Labs is an applied AI research group focused on open source models and the infrastructure that keeps them improving in production. The company has raised funding from Khosla Ventures and Insight Partners, with participation from M12, NEA, Valor Equity Partners, and several notable angels including former GitHub and Docker leaders and a co founder of Weights and Biases. Their GLiNER model family has seen wide adoption across the industry, which helps explain the push toward agent driven post training at scale.

Where this goes next

Every enterprise wants its own model, tuned to its own data. Very few have the internal bench to build and maintain it. What coding agents did for software productivity, Fastino is now attempting for specialized open weight language models. I think that matters. The path to trustworthy AI in sensitive domains is not a single frontier system. It is a set of right sized models that can live beside your data, learn from it responsibly, and answer the questions you actually have.