Two specialized open weight models for regulated industries, fine-tuned on NVIDIA Nemotron 3.5 Lightning by the Fastino Fine-Tuning Agent.

We are excited to announce a collaboration between Fastino Labs and NVIDIA to bring specialized open weight models to finance and healthcare.

Working with NVIDIA's Nemotron team, we've developed Fastino-Nemotron-3.5-Lightning-Finance and Fastino-Nemotron-3.5-Lightning-Healthcare, two specialized open weight models available on Hugging Face today, alongside the launch of NVIDIA Nemotron 3.5 Lightning.

Fastino-Nemotron-3.5-Lightning-Finance lifts FinQA accuracy of the base model from 15.86% to 59.23%, with gains across five financial reasoning benchmarks. Fastino-Nemotron-3.5-Lightning-Healthcare 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.

Both models were fine-tuned entirely by the Fastino Fine-Tuning Agent, an autonomous agent which we are also releasing today in private preview. Given a natural language prompt, the autoresearch agent ran the full post-training loop for them: no human created the training data or ran any fine-tuning experiments.

Available today

We are releasing:

• Two open weight specialized models. Fastino-Nemotron-3.5-Lightning-Finance and Fastino-Nemotron-3.5-Lightning-Healthcare. Both models are available today on Hugging Face under the Apache 2.0 license.

• The Fastino Fine-Tuning Agent, in private preview. The agent that fine-tuned both models, so your team can fine-tune Nemotron 3.5 Lightning and other open models for your own domains on your own data. Explore the agent sandbox at fastino.ai/agent or request full access at fastino.ai/waitlist.

The Fastino Fine-Tuning Agent produced both models in less than 10 hours, building upon the work of specialized prior runs with no human-in-the-loop. Given a natural language prompt, the agent determined the base model's failure patterns, researched the task and established state-of-the-art, sourced and curated its own training data, searched and experimented across multiple training configurations in parallel, ran its own contamination and transfer tests, and delivered the most accurate checkpoint along with the full training data and a report on all experiments run.

What coding agents did for software development, the Fastino Fine-Tuning Agent does the same for specialized open weight models. The user defines the problem and the agent runs the full post-training loop, from research to data curation through evaluation. Because they're no longer limited by how many experiments they can run manually, teams are free to focus on higher-value work.

What is Nemotron 3.5 Lightning?

Nemotron 3.5 Lightning is a 30B-parameter mixture-of-experts model with 3B active parameters, optimized for long-running agentic AI workloads. NVIDIA Nemotron models are fully open across weights, datasets, and training recipes. That openness makes agentic post-training possible and effective.

We started with finance and healthcare, two domains where privacy and governance restrict what teams can build with AI. The data that most improves a model is proprietary and tightly controlled, whether deal analyses, internal financial metrics, or clinical notes. open weight models, fine-tuned and run in your organization's environment, deliver performance while preserving data privacy and model sovereignty.

Both are also domains with real headroom. The base model already carries substantial financial and clinical knowledge. What post-training adds is the precision to apply it to the exact domain and task.

Behind our fine-tuned models: the Fastino Fine-Tuning Agent

The Fastino Fine-Tuning Agent is an auto research agent built on open research our team published earlier this year, and it has advanced considerably since.

Give the agent a task description in plain language and it runs the fine-tuning pipeline end to end: researching the task and current state-of-the-art, finding and curating training data, building an evaluation set before it trains, running multiple training configurations in parallel, recovering from failed runs, testing its own work for contamination, and shipping only the highest accuracy checkpoints. No human needs to generate training data. No human needs to run fine-tuning experiments.

Dhruv Atreja, Fastino Labs' Head of Agent Research and lead author on the agent paper, built the agent after watching how our own researchers work:

"Post-training a model is a guided search, so it's as much a process as it is a skill. It's a combination of a training recipe, your search parameters, and a guided dataset. If you give that to an agent as a Monte Carlo graph search, coding agents are very effective at iterating on that search policy."

We collaborated with the NVIDIA team on two model-specific fine-tuning recipes, which gave the agent a head start on both models. The agent can also work without a recipe, deriving its own approach for a model and domain it hasn't encountered before.

Searching those combinations of data, configuration, and supervision format is what used to take a researcher or ML team weeks. With the Fastino agent, fine-tuning a state-of-the-art open weight model can be achieved in less than a day.

Benchmark results

Fastino-Nemotron-3.5-Lightning-Finance

Fastino-Nemotron-3.5-Lightning-Finance is built to perform well across a variety of financial tasks. Across five financial benchmarks, it improves substantially over the base model.

These five benchmarks span a wide range of tasks across the finance domain, from numerical reasoning to extraction to open-ended financial research. FinQA and TAT-QA both test numerical reasoning over financial documents, the former over real SEC filings and the latter over hybrid tables and text. SEC-Num tests numerical span extraction from SEC earnings reports, while FinEntity measures entity-level recognition in financial text. Broader still is BizFinBench, which spans numerical calculation, reasoning, information extraction, prediction, and knowledge-based QA. Seeing gains this consistent across benchmarks this varied is a notable result for a single fine-tuned model.

FinQA measures execution accuracy on multi-step numerical reasoning over real SEC filings, and it produced the largest single gain of the finance release. The base model scores 15.86%, low enough that it cannot be used for this task. After fine-tuning, Fastino-Nemotron-3.5-Lightning-Finance reaches 59.23% on the test split, a gain of 43.37 points.

That score places a model with 3 billion active parameters in the range of general-purpose open weight models that activate ten times more or run fully dense. DeepSeek R1, a 671B-parameter model activating 37 billion parameters per token, scores 65.1%. Llama 3.3 70B, a dense model, scores 68.2%. The best open weight score we found, Kimi K2.6 at 76.0%, activates 32 billion parameters per token. Every model in the comparison set activates at least ten times the parameters ours does, and the gap to the strongest of them is around 11 points.

Transfer to unseen benchmarks

The clearest evidence that the model learned reusable financial reasoning, rather than just fitting the benchmarks it trained on, is its performance related tasks the model was not specifically trained for.

On ConvFinQA, a conversational financial-reasoning benchmark the model was not directly improved to beat, performance improved by over 40 points. It also improved on FiQA, with no regression. Gains that appear on held-out benchmarks come from reasoning the model carries over on its own, which is what distinguishes genuine capability from memorized benchmark formats.

Fastino-Nemotron-3.5-Lightning-Healthcare

Fastino-Nemotron-3.5-Lightning-Healthcare is built for clinical conversation quality, medical reasoning, and extracting medical concepts from unstructured text. Across eight healthcare benchmarks and two transfer evaluations, the fine-tuned model improves over the base model on every task, with gains ranging from 4.00 to 24.59 points. Three of the evaluations were blind, held out from the training loop entirely, and two were run on tasks the model was not explicitly taught during training. All of it came from a single autonomous run: the agent researched the tasks, built the training data, and searched the recipe space without a human running any experiment.

These eight benchmarks span a wide range of tasks across the healthcare domain, from agentic EHR workflows to clinical conversation to medical reasoning to concept extraction.

• MedAgentBench evaluates agents on physician-written clinical tasks inside a FHIR-compliant electronic health record environment, covering workflows such as retrieving patient data and placing orders.

• HealthAdminBench tests agents on healthcare administration workflows, including prior authorization, appeals and denials management, and durable medical equipment orders, across realistic EHR and payer-portal environments.

• BC5CDR measures extraction of chemical and disease mentions from biomedical text.

• MedMentions ST21pv measures concept extraction across a broad set of UMLS semantic types.

• MEDEC tests detecting and correcting errors in clinical notes, reported here as separate flag and sentence accuracies.

• MedCalc requires extracting patient values and applying the correct clinical calculator.

• PubMedQA tests yes/no/maybe reasoning on biomedical research abstracts.

• HealthBench grades multi-turn health conversations against physician-written rubrics.

On our full internal MedCalc-Bench Verified evaluation, spanning all 1,100 cases, Fastino-Nemotron-3.5-Lightning-Healthcare improves over the base model by 5.09 points (49.09% to 54.18%). On the complete 574-case MEDEC evaluation, it improves across all three measured behaviors, lifting flag accuracy from 53.66% to 64.98% and sentence accuracy from 48.08% to 62.89%. On flag accuracy our model beats Claude Opus 4.6 by 6.94 points and comes within 0.06 points of matching Muse Spark, demonstrating remarkable performance by an open weights model of this size.

Transfer to unseen benchmarks

As with the finance model, the clearest evidence of reusable capability is performance on healthcare benchmarks the model never saw during training.

BioRED tests biomedical entity and relation extraction on a dataset the model never trained on; the gain there shows that the extraction capability built on BC5CDR's chemical and disease mentions carries over to a harder task with new entity types and document-level relations. EkaCare tests selecting and applying the correct medical calculator to a clinical scenario, drawn from a separate bilingual Indian-healthcare dataset. The model improved substantially on both, with no regression on either.

Get started

Today, we release both Fastino-Nemotron-3.5-Lightning-Finance and Fastino-Nemotron-3.5-Lightning-Healthcare under the Apache 2.0 license.

To learn more about how our agent fine-tuned the models and best practices for fine-tuning Nemotron models, please refer to our technical blog post. For more information about the Fastino Fine-Tuning Agent, including preview access, refer to our website.

• Finance model: https://huggingface.co/fastino/Fastino-Nemotron-3.5-Lightning-Finance

• Healthcare model: https://huggingface.co/fastino/Fastino-Nemotron-3.5-Lightning-Healthcare

• License: Apache 2.0

• Base model: NVIDIA Nemotron 3.5 Lightning

• Try the Fastino Fine-Tuning Agent: Agent sandbox

• Fastino Fine-Tuning Agent: Request private preview access

• Research: Read the paper