Fastino Nemotron 3.5 Super VL Finance

Fastino-Nemotron-3.5-Super-VL-Finance is a 120B-parameter, 12B-active hybrid Mamba-Transformer mixture-of-experts model specialized for financial reasoning, extraction, and research. The repository contains a ready-to-run merged checkpoint: the language backbone of the NVIDIA Nemotron 3.5 Super VL early-access checkpoint plus Fastino's combined-finance LoRA adapter. Users do not need to obtain or attach a separate LoRA.

• Developed by: Fastino Labs, in collaboration with NVIDIA
• Base checkpoint: NVIDIA Nemotron 3.5 Super VL, early-access checkpoint (language backbone)
• Post-training checkpoint: Fastino-Finance
• License: Apache 2.0
• Language: English
• Modalities: Text (the vision encoder of the VL base is not included)
• Fine-Tuning Agent: Agent sandbox
• Research: Fastino Fine-Tuning Agent

What it is designed for

The model targets financial document reasoning, numerical question answering over filings and tables, numeric span extraction, financial entity recognition, conversational analysis, and source-grounded financial research. The evaluation suite includes FinQA, TAT-QA, SEC-Num, FinEntity, BizFinBench, ConvFinQA, FiQA, and EDGARCalcQA.

Quickstart

The published weights are BF16 and require about 247 GB before runtime overhead. Tensor parallelism across four 80 GB or larger GPUs (or two B200s) is recommended.

pip install vllm  # a release with Nemotron 3.5 Super support

from vllm import LLM, SamplingParams

model_id = "fastino/Fastino-Nemotron-3.5-Super-VL-Finance"

llm = LLM(
    model=model_id,
    trust_remote_code=True,
    dtype="bfloat16",
    tensor_parallel_size=4,
    max_model_len=8192,
)

outputs = llm.generate(
    ["Check this financial calculation and explain the result: ..."],
    SamplingParams(temperature=0.0, max_tokens=512),
)
print(outputs[0].outputs[0].text)

Post-training recipe

The data recipe was discovered by the Fastino Fine-Tuning Agent on NVIDIA Nemotron 3.5 Lightning, where it autonomously built evaluation sets, curated data, explored training mixtures and hyperparameters, evaluated transfer, and selected the winning mixture. For Super we ran no new search: the same mixture was trained once on the Super checkpoint through the Fastino Training API.

The adapter was trained on 13,698 de-duplicated examples covering:

• financial document calculation and executable reasoning;
• hybrid text-and-table question answering;
• business-finance reasoning across calculation, extraction, temporal reasoning, prediction, and knowledge tasks;
• numerical span extraction from SEC disclosures;
• financial entity extraction; and
• source-grounded financial research trajectories.

It was trained from the base checkpoint for two epochs with LoRA rank 32 on attention, Mamba, latent-projection, and shared-expert weights (75.2M trainable parameters), learning rate 2e-4, batch size 16, and sequence packing disabled. Evaluation examples and labels were excluded from training.

Benchmark evaluation

Base and fine-tuned scores below use the same early-access checkpoint, inputs, prompts, decoding settings, inference route, and evaluator for each row.

In-domain benchmarks

Transfer to unseen benchmarks

Performance on related tasks the model was not explicitly trained for.

Evaluation protocol

Every benchmark uses a held-out set fixed before training and the same prompts and scorers as our Nemotron 3.5 Lightning study. Score-bearing comparisons were accepted only when base and candidate shared the same examples, prompt construction, decoding parameters, serving route, evaluator identity, and aggregation. Training mixtures were hashed and de-duplicated, and benchmark evaluation rows were excluded from training. Results are from a single training run.

Limitations

This is a specialized model whose capabilities are best characterized by the tasks above. Performance outside these tasks has not been comprehensively evaluated. Outputs are not a substitute for professional financial advice; high-stakes use requires independent validation and qualified human oversight.

Fine-tuning

To fine-tune or serve models like this one without managing GPUs, see the Fastino API, which offers hosted training and inference.

Citation

@misc{atreja2026pioneeragentcontinualimprovement,
      title={Pioneer Agent: Continual Improvement of Small Language Models in Production},
      author={Dhruv Atreja and Julia White and Nikhil Nayak and Kelton Zhang and Henrijs Princis and George Hurn-Maloney and Ash Lewis and Urchade Zaratiana},
      year={2026},
      eprint={2604.09791},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2604.09791},
}

License

This model is licensed under the Apache License 2.0.

Links

• Fastino Labs
• Supporting research
• NVIDIA Nemotron 3.5 Super VL base model