AI That Trains Itself: Fastino's Agent Unlocks Finance & Healthcare

PALO ALTO, CA – August 11, 2026 – The chasm between the promise of artificial intelligence and its practical, secure deployment within regulated industries has long been defined by the high costs of expertise, time, and data risk. Today, that chasm may have significantly narrowed. Fastino Labs, an applied AI research firm, announced the release of two open-weight models for finance and healthcare that were not trained by a human team, but by a fully autonomous AI agent in less than a day.

This development, centered on the new Fastino Fine-Tuning Agent, represents more than just an incremental improvement in model performance. It signals a potential paradigm shift in how specialized AI is created and deployed, moving the goalposts from a months-long, human-intensive research project to an overnight, automated process. By building on NVIDIA's open-weight Nemotron 3.5 Lightning model, Fastino is providing a compelling case study in how autonomous systems can tackle one of the biggest bottlenecks in enterprise AI adoption.

The Automation of Expertise

The most significant part of Fastino's announcement isn't the models themselves, but how they were made. The Fastino Fine-Tuning Agent, released today in a private preview, is an autonomous system that takes a simple, plain-language goal—in this case, building high-performing finance and healthcare models—and executes the entire complex workflow required to achieve it.

Traditionally, post-training or fine-tuning a large language model is an arcane art. It requires a dedicated team of scarce, expensive ML engineers to spend weeks or even months on a guided-but-manual process of data curation, experiment design, parameter tuning, and failure analysis. The Fastino agent automates this entire pipeline. According to the company, it researches the task, sources and curates training data, runs multiple experiments in parallel, and delivers the best-performing model—all without a human engineer running a single experiment. The result: two domain-specific models produced in under ten hours.

"Post-training a model is as much a guided search process as it is a skill, combining a training recipe, your search parameters, and a curated dataset," explained Dhruv Atreja, Head of Agent Research at Fastino Labs. "If you give that to an agent as a Monte Carlo Graph Search, it becomes very effective at iterating on that search policy." This reframes the challenge from a human skill problem to a computational search problem, one that machines are uniquely suited to solve at scale.

The implications are profound. For enterprises, the hidden costs of building a specialized AI team—recruitment, salaries, infrastructure, and the opportunity cost of long development cycles—have been a prohibitive barrier. By automating this workflow, the agent effectively democratizes access to bespoke AI. "Every enterprise wants a model trained on its own data, and very few have a post-training team to build it," said George Hurn-Maloney, co-founder of Fastino Labs. "These two models are proof that you no longer need one."

Unlocking Data in Fortified Industries

Nowhere is the need for specialized, secure AI more acute than in finance and healthcare. These sectors are sitting on mountains of high-value proprietary data—clinical notes, patient histories, financial filings, market analysis—that could be used to train incredibly powerful AI tools. Yet, this data is rightly locked down by stringent privacy and governance regulations like HIPAA and financial compliance standards.

This has created an unacceptable choice for many firms: use a general-purpose frontier model from a major provider, which often means sending sensitive data to a third-party API and losing control, or do nothing and fall behind. Fastino's approach directly addresses this dilemma. The new models are open-weight, released under the permissive Apache 2.0 license. This means an organization can download the model, fine-tune it on its own proprietary data using the agent, and run it entirely within its own secure, on-premise infrastructure. The sensitive data never has to leave the building.

The performance gains demonstrate this is not a compromise. The Fastino-Nemotron-3.5-Lightning-Finance model saw its accuracy on the FinQA benchmark—a difficult test of numerical reasoning over real SEC filings—skyrocket from a nearly unusable 15.86% in the base model to a highly competitive 59.23%. This single improvement transforms the model from a novelty into a potentially powerful tool for financial analysts.

Similarly, the healthcare model posted significant gains across eight different medical benchmarks. On the MEDEC benchmark, which evaluates a model's ability to identify errors in medical text, the new model's accuracy jumped over 11 points, placing it on par with or ahead of much larger, closed-source models. This level of performance, combined with the ability to maintain full data sovereignty, presents a clear, actionable path for healthcare organizations to begin deploying AI for tasks like clinical documentation and medical reasoning.

The Power of an Open Ecosystem

This breakthrough was not created in a vacuum. It stands on the shoulders of an increasingly vibrant open-source AI ecosystem, with NVIDIA's strategic decisions playing a pivotal role. The entire process was made possible because NVIDIA chose to release its Nemotron 3.5 Lightning model with open weights, datasets, and training recipes. A closed, black-box model gives an autonomous agent nothing to work with; an open model provides the foundational architecture and data for the agent to reason about and build upon.

"Open models give enterprises the ability to fine-tune on their own proprietary data and retain full control of their IP — a critical advantage in data-sensitive domains like finance and healthcare," said Erik Pounds, senior director of Enterprise AI at NVIDIA. This partnership highlights NVIDIA's broader strategy: by providing powerful, open foundational tools, it fosters an ecosystem of innovation where partners like Fastino can create specialized, high-value solutions. This, in turn, drives demand for the underlying hardware and software that powers the entire stack.

Fastino Labs, which has already seen its open-source GLiNER models downloaded over 30 million times, is now extending this philosophy to the very process of model creation. The agentic system they've built is a testament to how layers of open technology can compound, leading to capabilities that were unthinkable just a few years ago. What coding agents began doing for software development, the Fastino agent is now doing for the development of AI itself, marking a new chapter in the evolution of artificial intelligence.