Nvidia's Medical Fine-Tuning Cuts Nemotron 3.5 Transcription Error Rate by 45.8%

Nvidia's medical fine-tuning of its Nemotron 3.5 Transcribe model reduced the word error rate on medical recordings from 22.44% to 12.17%, a 45.8% improvement achieved using about 14 GPU hours on a single GB300 processor. The company is publishing the fine-tuning recipe and model weights to help developers adapt the tool to specialized vocabularies. Nemotron 3.5 Transcribe is a 30B parameter Mixture-of-Experts model supporting 40 languages and speaker identification, available for commercial use under the OpenMDW-1.1 license.