# Heidi fine‑tunes NVIDIA Nemotron 3 Ultra on Fireworks for clinical AI

> Heidi post‑trained the 550B Nemotron 3 Ultra model on Fireworks, achieving a 54% win rate in live clinician side‑by‑side tests versus their current frontier model.

Oossa · 2026-10-23 · https://oossa.com/en/heidi-fine-tunes-nvidia-nemotron-3-ultra-on-fireworks-for-clinical-ai

Heidi used Fireworks’ training API to supervised‑fine‑tune (SFT) and then apply reinforcement learning (RL) to NVIDIA’s Nemotron 3 Ultra, a 550‑billion‑parameter open model. After quantizing to NVIDIA’s NVFP4 precision, they deployed the checkpoint on dedicated GPUs. In live side‑by‑side clinician comparisons, the post‑trained model beat the existing production model 54% of the time, up from about 42% for the base checkpoint.

## The facts

- Base checkpoint win rate: ~42%; post‑trained win rate: ~54%
- Model size: 550 B total parameters, 55 B active per token

## Why it matters

Clinicians can rely on a model they control, allowing updates and on‑premises deployment for hospitals with strict data‑residency rules.

## Sources & references

1. [Heidi trains and deploys NVIDIA Nemotron 3 Ultra on Fireworks](https://fireworks.ai/customers/heidi-trains-and-deploys-nvidia-nemotron-3-ultra-on-fireworks/) – Fireworks AI

Last updated: 2026-10-23