FROM THE CREATOR OF GLINER
Describe the task and let the Fastino agent generates high-quality data, runs full or LoRA training, and automated evals. All from your coding agent.
HOW THE AGENT WORKS
01
Research the task and objectives
The agent starts researching and scoping the project based on your supplied tasks and objectives.
02
Source or generate synthetic data
Once training plan is approved, the agent sources, generate high-quality synthetic data, or curates supplied datasets.
03
Run experiments & evaluations
The agent can perform LoRA sweeps or full fine-tunes, and automated evals with provenance receipts.
04
Ship the checkpoint
Deploy the endpoint for inference, or push the checkpoint to your Hugging Face space.
The agent cli
Agent to agent. Claude Code, Codex, or Cursor hands the task to the Fastino agent, which trains your model in hours and returns the checkpoint, training data, and eval report.
claude code · ~/acme/financial-qa
› use the Fastino agent to fine-tune a financial QA model over our 10-K filings
I will hand this to the Fastino agent over A2A and stream its run back here.
A2A call · fastino.train
task: finance_qa
base_model: Fastino-Nemotron-3.5-Lightning-Finance
corpus: ./data/10k_filings/*.pdf
eval: ./evals/finance_qa_holdout.jsonl
budget: lora_then_full · max_hours: 8
fastino · run finqa-9d2e accepted · streaming
Parsed and normalized the 10-K corpus
214 filings · 1.8M tokens · source citations preserved
Generated and validated training data
32k grounded Q&A pairs filtered for answer and citation quality
Ran 6 LoRA sweeps, promoted the winner to a full fine-tune
grounded answer accuracy 0.89 · citation F1 0.93 · +18 over zero-shot
Deployed and pushed the checkpoint
POST /v1/finance-qa/answer · fastino/Nemotron-Finance-10K-QA
run complete in 6h 48m. checkpoint, training data, and eval report written to ./fastino/finqa-9d2e/
› now connect it to the financial analysis app
When used to build specialized models for finance and healthcare, Fastino Agent improved performance vs. Nemotron 3.5 Lightning base model.
Fastino-Nemotron-3.5-Lightning-Finance
Executable financial programs, table arithmetic, and numerical extraction: six finance skills in a single rank-32 adapter.
FinQA
15.9→ 59.2%
TAT-QA (F1)
19.0 → 56.6%
SEC-Num
79.7 → 87.6%
Outperforms Nemotron 3 Ultra on BigFinanceBench while activating 3B parameters per token.
Open on Hugging Face
Fastino-Nemotron-3.5-Lightning-Healthcare
Physician-graded conversation, biomedical reasoning, clinical calculation, error correction, and concept extraction: six clinical skills, one adapter.
HealthBench
49.7 → 56.2%
PubMedQA
59.0 → 65.0%
MEDEC
50.9 → 63.9%
Within 0.1 points of GPT-5.5 Instant on HealthBench, at a fraction of the active parameters.
Open on Hugging Face
Fine-tune Fastino-Nemotron models on your own data
The autonomous agent that built the Fastino-Nemotron-3.5-Lightning Finance & Healthcare models is now available in private preview.
Get access