Fastino Labs is betting that enterprises in highly regulated industries need more than a capable general-purpose AI model: they need models they can run, inspect and fine-tune on proprietary data. The applied AI research company has released two open-weight models for finance and healthcare, post-trained on NVIDIA Nemotron 3.5 Lightning by an autonomous fine-tuning agent rather than a conventional human research team.
Fastino Uses Autonomous AI Research to Build Specialized Finance and Healthcare Models
The next phase of enterprise AI may depend less on finding one model that can do everything and more on making smaller open models exceptionally good at specific jobs.
Fastino Labs is pursuing that strategy with two new open-weight models designed for data-sensitive industries: Fastino-Nemotron-3.5-Lightning-Finance and Fastino-Nemotron-3.5-Lightning-Healthcare.
Both models are based on NVIDIA Nemotron 3.5 Lightning, a 30-billion-parameter mixture-of-experts model, and were post-trained using Fastino’s autonomous Fine-Tuning Agent.
The unusual part is how the models were developed.
Fastino says neither model was post-trained by a person. Its autonomous agent handled the research, dataset selection, evaluation design, training experiments and model selection, producing each model in less than 10 hours.
The company is positioning the release as a demonstration of two trends converging: open-weight enterprise AI and agentic AI research automation.
Open models address a critical enterprise constraint
Finance and healthcare are among the most attractive applications for AI—and among the hardest places to deploy it.
Financial institutions have proprietary filings, transaction histories, research and internal metrics. Healthcare organizations hold clinical notes, documentation and other sensitive information. Sending that data to a closed third-party model can create privacy, compliance, intellectual-property and data-governance concerns.
Open-weight models offer a different architecture.
An enterprise can deploy the model in its own environment and fine-tune it using proprietary information without necessarily sending that underlying data to an external model provider.
That does not eliminate regulatory obligations or security risks. But it gives organizations significantly more control over where models run and how proprietary data is incorporated.
This is one reason the open-model ecosystem has become strategically important alongside closed systems from OpenAI, Google, Anthropic and Microsoft.
Fastino’s argument is that specialized open models could make that approach more practical for industries where general-purpose AI often lacks the necessary domain precision.
The Finance model targets numbers, filings and financial research
Fastino-Nemotron-3.5-Lightning-Finance is aimed at financial-document analysis, numerical reasoning, research and summarization.
The company’s reported benchmark results show substantial improvements over the underlying Nemotron model.
On FinQA, which tests numerical reasoning using real financial filings, Fastino reports that execution accuracy increased from 15.86% to 59.23%, a gain of 43.37 percentage points.
The model also improved on TAT-QA, SEC-Num, FinEntity and BizFinBench, according to Fastino’s evaluation.
Those improvements matter because financial AI is often less about generating fluent text and more about getting the arithmetic and relationships between numbers right.
A model summarizing an annual report can be useful. A model that can correctly reason through a table, calculate a financial metric and explain the result is much closer to a production financial-research tool.
Fastino says the Finance model also showed gains on related tasks it was not specifically trained for, which the company argues indicates broader domain learning rather than benchmark-specific optimization.
That claim will need independent validation as enterprise teams assess the models, particularly because benchmark performance does not automatically translate into reliable performance on proprietary financial documents.
Healthcare presents an even higher bar
The healthcare model targets clinical conversations, medical reasoning, documentation and extraction of medical concepts from unstructured text.
Fastino reports improvements across several medical benchmarks, including MEDEC, MedCalc-Bench, MedMentions and BC5CDR.
On the full 1,100-case MedCalc-Bench Verified evaluation, the model reportedly improved from 49.09% to 54.18%.
On MEDEC, Fastino reports an increase in flag accuracy from 53.66% to 64.98%.
The results are notable for an open-weight model, but healthcare deployment demands considerably more than benchmark accuracy.
Clinical AI must contend with patient privacy, hallucinations, clinical validation, regulatory requirements and the consequences of incorrect recommendations.
For that reason, the more immediately practical use cases may be documentation, extraction, summarization and workflow assistance rather than autonomous diagnosis or treatment decisions.
The advantage of an open model is that healthcare organizations can potentially adapt it to their own terminology, workflows and documentation standards while maintaining greater control over the environment in which the model operates.
The more interesting announcement may be the agent
The models themselves are only half of Fastino’s announcement.
The company’s Fine-Tuning Agent is designed to automate what is traditionally a highly specialized machine-learning workflow.
Given a task description, Fastino says the agent researches existing approaches, identifies and curates training data, creates an evaluation set, runs multiple training and data experiments, recovers from failed jobs, checks for contamination and selects the strongest checkpoint.
That process is fundamentally different from simply running a predetermined hyperparameter sweep.
Fastino describes it as a search through entire post-training pipelines, where the agent uses the results of previous experiments to determine what to try next.
The comparison the company makes is straightforward: coding agents automated parts of software development; its Fine-Tuning Agent is intended to automate parts of model development.
If that approach scales, it could reduce one of the biggest barriers to enterprise model customization: the shortage of researchers capable of conducting sophisticated post-training work.
Nemotron’s openness is part of the equation
The agent also depends on the openness of the underlying model ecosystem.
NVIDIA’s Nemotron strategy makes model weights, datasets and training recipes available, giving researchers more components to experiment with than they would have when working with a closed commercial model.
NVIDIA is positioning Nemotron 3.5 Lightning for long-running agentic workloads and specialized tasks within larger multi-agent systems.
Fastino’s work demonstrates another potential use: taking an already capable open model and using automated post-training to specialize it for a particular industry.
That architecture could become increasingly common.
Rather than choosing between a giant proprietary model and a generic open model, enterprises may build systems containing several specialized models, each optimized for a specific workflow.
Enterprise AI moves toward customization
The implications extend beyond finance and healthcare.
Large enterprises increasingly want models adapted to internal terminology, proprietary workflows and specialized datasets. But traditional fine-tuning requires machine-learning expertise, compute resources and weeks of experimentation.
Automating that process could change the economics.
The key question is whether autonomous post-training can produce models that remain reliable outside benchmarks and across messy real-world enterprise data.
That will require independent evaluations, transparent training datasets, reproducible recipes and careful testing for contamination and safety.
Fastino’s release is therefore less a finished enterprise solution than an indication of where AI infrastructure may be heading.
The frontier is moving from “Which model should we use?” toward “How quickly can we build the right model for this job?”
For finance and healthcare teams that cannot freely move proprietary data into closed AI systems, that distinction could become one of the most important developments in enterprise AI adoption.
Market Landscape
Fastino’s announcement sits within a rapidly developing open-model ecosystem that includes NVIDIA Nemotron, Meta Llama, Mistral AI, Google Gemma and other open-weight model families.
Several trends are converging:
- Domain-specific AI: Enterprises increasingly want models optimized for finance, healthcare, legal, cybersecurity and other specialized workflows.
- Private AI infrastructure: Sensitive organizations are seeking deployment options that provide greater control over proprietary data.
- Agentic model development: AI agents are beginning to automate not only application development but parts of the machine-learning research process.
- Post-training as a competitive layer: Model specialization can become a differentiator even when organizations start with the same foundation model.
- Multi-model architectures: Enterprises may increasingly combine general-purpose models with smaller specialized models rather than relying on one system for every workload.
The competitive landscape is moving quickly. Closed frontier models remain attractive for breadth and general reasoning, while open-weight models offer organizations greater control and customization.
Fastino is effectively arguing that automated post-training can narrow the performance gap without requiring every enterprise to build its own AI research organization.
Top Insights
- Fastino released open-weight finance and healthcare models based on NVIDIA Nemotron 3.5 Lightning, targeting sensitive workloads requiring greater control over proprietary data.
- An autonomous Fine-Tuning Agent developed both models in under 10 hours, automating dataset selection, evaluation, training experiments and checkpoint selection.
- The Finance model reportedly raised FinQA execution accuracy from 15.86% to 59.23%, demonstrating substantial gains in financial numerical reasoning.
- The Healthcare model improved across multiple medical benchmarks, including MedCalc-Bench and MEDEC, although production clinical use requires extensive validation.
- Automated post-training could lower the expertise barrier for enterprises seeking specialized open models without maintaining dedicated AI research teams.
Get in touch with our fintech expert
