# NeMo NVIDIA NeMo is an open-source, cloud-native generative AI framework built on PyTorch. Designed for researchers and developers working on Large Language Models (LLMs), Multimodal models, and Speech AI (ASR and TTS). Licensed under Apache 2.0. ## Key features - Scalable training across thousands of GPUs via PyTorch Lightning - Distributed training: Tensor Parallelism, Pipeline Parallelism, FSDP, Mixture-of-Experts (MoE) - Mixed precision training (BFloat16, FP8) - Parameter-efficient fine-tuning: LoRA, P-Tuning, Adapters - NeMo 2.0: Python-based configuration replacing YAML, modular abstractions ## Domains - **LLMs**: Training and fine-tuning transformer-based language models - **Speech AI**: Automatic Speech Recognition (ASR) and Text-to-Speech (TTS), including the Parakeet model family - **Multimodal**: Vision-language and cross-modal architectures - **Computer Vision**: Image generation and processing ## Ecosystem - **NeMo Guardrails**: Programmable safety layer (topical relevance, PII redaction, jailbreak prevention) - **NeMo Curator**: Data curation library, including optimized video processing - **NeMo-Run**: Experiment management for large-scale training - **BioNeMo**: Specialized platform for AI-driven drug discovery ## References - Website: https://www.nvidia.com/en-us/ai-data-science/products/nemo/ - Source code: https://github.com/NVIDIA-NeMo/NeMo - Documentation: https://docs.nvidia.com/nemo-framework/user-guide/latest/overview.html - Pre-trained models: https://huggingface.co/nvidia ## Related - [[Parakeet V3]] - [[Speech-to-Text (STT)]] - [[Text-to-Speech (TTS)]]