Open Source
Explore the latest AI open-source projects from GitHub and HuggingFace.
Explore the latest AI open-source projects from GitHub and HuggingFace.
Megatron-LM is NVIDIA's GPU-optimized library for distributed training of large transformer models at scale. It provides composable building blocks through Megatron Core, supporting advanced parallelism strategies including Tensor Parallel, Pipeline Parallel, Data Parallel, Expert Parallel, and Context Parallel. The library achieves up to 47% Model FLOP Utilization on H100 clusters and scales from 2B to 462B parameters across thousands of GPUs. Recent updates include Dynamic Context Parallelism with 1.48x speedup for variable-length sequences, Megatron Bridge for Hugging Face checkpoint interoperability, and expanded MoE support for DeepSeek-V3 and Qwen3 architectures.
huggingface
The de-facto standard ML framework with 159k+ stars — centralized model definitions for Llama, Qwen, Whisper, DINO, and 1M+ pretrained checkpoints.
langchain-ai
The dominant open-source agent engineering platform for building LLM-powered applications with 132k+ GitHub stars, supporting chains, agents, RAG, and multi-provider LLM integration.
open-webui
Self-hosted AI platform with 127K+ stars supporting Ollama, OpenAI APIs, built-in RAG with 9 vector databases, and enterprise access controls.
deepseek-ai
671B MoE open-weight LLM achieving frontier-level performance on reasoning and coding benchmarks