Fastino Labs Releases Two Specialized Open-Weight Models for Finance and Healthcare, Post-Trained on NVIDIA Nemotron 3.5 Lightning Entirely using an Agent

Fastino-Nemotron-3.5-Lightning-Finance and Fastino-Nemotron-3.5-Lightning-Healthcare, post-trained on Nemotron 3.5 Lightning by the Fastino Fine-Tuning Agent, are available on Hugging Face today under the Apache 2.0 license. Both models outperform the Nemotron 3.5 Lightning base model across a variety of finance and healthcare benchmarks, showing off both the ease of fine-tuning Nemotron and the power of doing it with the Fastino agent. The agent that post-trained them is also available today in private preview.

PALO ALTO, Calif., Aug. 11, 2026 /PRNewswire/ -- Fastino Labs, the applied AI research lab and creator of the GLiNER model family, today announced specialised open weight models for finance and healthcare post-trained on NVIDIA Nemotron 3.5 Lightning.

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 today.

Neither model was post-trained by a person. Both were post-trained end-to-end by the Fastino Fine-Tuning Agent, a fully autonomous autoresearch agent, which the company is also releasing today in private preview. The agent produced each model in less than ten hours, work a post-training team would have likely spent weeks on.

The finance model lifts FinQA accuracy of the base model from 15.86% to 59.23%, with gains across five financial reasoning benchmarks while the healthcare model delivers verified gains on MEDEC and MedCalc-Bench and lifts performance on a total of eight medical benchmarks. Both models showed gains that transferred to related tasks they were never trained for, showing that the gains reflect genuine generalized domain capabilities rather than fitting to specific benchmarks.

Models that excel in data sensitive domains

Open models specialized for finance and healthcare remain scarce, and many are years old. The public FinQA leaderboard tracks three models, all proprietary. HealthBench's leaderboard is dominated by closed models. That leaves teams in the two domains with an unacceptable choice: using a general-purpose frontier model they cannot run on their own data, or doing nothing.

Data privacy and governance have been holding LLM adoption back for high ROI use cases. The data that would most improve a model, whether filings, clinical notes, or internal financial metrics, is proprietary and tightly governed, and can rarely leave the environment it lives in. An open weight model that a team can fine-tune and run on its own infrastructure is how that team gets domain accuracy without giving up control of the data that produces it.

Both domains also have real headroom. Nemotron 3.5 Lightning already carries substantial financial and clinical knowledge. What post-training adds is the precision to improve it on an organization's proprietary data for their specific use case.

Post-training without a research team

The Fastino Fine-Tuning Agent is a fully autonomous autoresearch agent that takes a task description in plain language and runs the entire fine-tuning pipeline: researching the task and the current state-of-the-art, finding and curating training data, building an evaluation set before training begins, running multiple training configurations and data experiments in parallel, recovering from failed runs, testing its own work for data contamination, and delivering the highest accuracy checkpoint. In this case, the agent was tasked with building two models that would perform well across a wide range of finance and healthcare tasks.

No Fastino research engineer had to select or generate the training data, or run a single fine-tuning experiment.

Any training run comes down to three interlocking decisions: what data to train on, how to configure the run, and what supervision format to use. Those decisions cannot be separated, which is why a fixed parameter grid does not work and fine-tuning by hand takes weeks. The agent searches over whole pipelines instead, and each attempt is a reasoned response to the last one rather than a point on a grid.

"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," said 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."

What coding agents did for software, the Fastino Fine-Tuning Agent does for specialized open-weight models. It is built on research the company published earlier this year and has grown considerably since: it now runs far longer per job, carries what it learned on previous runs into new ones, and handles model architectures the original system could not, which is what made post-training a newly released Nemotron model possible. Fastino will be launching the agent for full-release in the coming weeks with support for a variety of open weight models

"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. In less than a day, an agent did the work a post-training team would have spent weeks on. It's extremely exciting to think how agentic systems can advance the field of language model research."

"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, Enterprise AI, NVIDIA. "Fastino's work shows how developers can use NVIDIA Nemotron 3.5 Lightning to build specialized models to deploy on their own or within larger systems of models to deliver domain accuracy at a fraction of the cost."

Built on open weights, datasets, and recipes

NVIDIA is expanding its Nemotron 3 model family with Nemotron 3.5 Lightning, the highest-efficiency model with leading accuracy in its class for long-running agentic AI workloads, including personal assistant and knowledge work. Built for specialized tasks within larger multi-agent systems, Nemotron 3.5 Lightning, a 30-billion-parameter mixture-of-experts model, helps create smarter and efficient agentic applications.

Nemotron is open across weights, datasets, and training recipes, and that openness is what makes agentic post-training possible. A closed model gives the agent nothing to search over. Every architectural choice the Nemotron team published is one the agent can reason about, and every dataset they released is one it can draw from.

Fastino-Nemotron-3.5-Lightning-Finance: numerical reasoning over financial documents

Fastino-Nemotron-3.5-Lightning-Finance is built for numerical reasoning, research, and summarization over financial concepts and documents such as filings, reports, and the dense tables inside them. Across six financial benchmarks spanning numerical reasoning, extraction, and open-ended financial research, it improves substantially over the Nemotron base model.

On FinQA, which tests multi-step numerical reasoning over real SEC filings, fine-tuning lifted the model from 15.86%, too low for practical use, to 59.23%, a gain of 43.37 points and the largest of the finance release. On a leaderboard led by frontier systems in the 55 to 67% range, a 30B open model with 3B active parameters clearing 40% is meaningful financial-research capability at a fraction of the size and cost.

Fastino-Nemotron-3.5-Healthcare: clinical conversation and documentation

Fastino-Nemotron-3.5-Healthcare is built for clinical conversation quality, medical reasoning, summarizing notes and discharge documentation, and medical concept extraction from unstructured text. On HealthBench, which grades multi-turn clinical conversations against rubrics written by 262 physicians, the fine-tuned model improves the base model from 40.62% to 48.17% across 700 held-out conversations. Across six healthcare benchmarks it improves significantly over the Nemotron base model.

On the full 1,100-case MedCalc-Bench Verified evaluation, Fastino-Nemotron-3.5-Lightning-Healthcare improves over the base model by 5.09 points (49.09% to 54.18%). On MEDEC, it improves across every measured behavior, lifting flag accuracy by 11.32 points to 64.98%, ahead of Claude Opus 4.6 and within 0.06 points of Muse Spark, a strong result for an open-weight model of this size.

Availability

Fastino-Nemotron-3.5-Lightning-Finance and Fastino-Nemotron-3.5-Lightning-Healthcare are available on Hugging Face today under the Apache 2.0 license.

The Fastino Fine-Tuning Agent is available in private preview. Teams can request access at Fastino.ai today.

Backed by leading investors

Fastino Labs is an applied AI research lab building open source models and the infrastructure to make them continuously better in production. Founded in 2024, Fastino Labs has raised $25 million in total funding across its pre-seed and seed rounds, led by Khosla Ventures and Insight Partners, with participation from M12 (Microsoft's venture fund), NEA, Valor Equity Partners, and angels including former GitHub CEO Thomas Dohmke, former Docker CEO Scott Johnston, and Weights & Biases co-founder Lukas Biewald.

Fastino's open source GLiNER model family has been downloaded more than 30 million times and is used in production by teams at leading Fortune 500 companies.

Learn more at fastino.ai.

SOURCE Fastino Labs