Open Source
Explore the latest AI open-source projects from GitHub and HuggingFace.
Explore the latest AI open-source projects from GitHub and HuggingFace.
Explore the latest AI open-source projects from GitHub and HuggingFace. Discover projects across categories like LLM, Vision, Audio, and more.
760 projects
omnigent-ai
Apache-2.0 meta-harness that puts one orchestration, policy and sandboxing layer over Claude Code, Codex, Cursor, Pi and custom agents — with sessions that follow you from terminal to phone.
openlake-project
Apache-2.0 Rust storage engine on io_uring that turns GPU nodes into a petabyte-scale KV cache pool — a drop-in vLLM connector reporting 66x faster time-to-first-token on cached 128K contexts.
microsoft
Microsoft's MIT-licensed 4B image-to-3D model using a field-free O-Voxel representation — handles open surfaces and non-manifold geometry, exports full PBR assets in ~3s at 512³.
OpenMOSS
Apache-2.0 0.1B-parameter multilingual TTS from OpenMOSS — 48 kHz stereo output, 20 languages, streaming on a 4-core CPU, and an ONNX build that runs on a single core of a MacBook Air M4.
Open-Less
MIT-licensed open-source voice input for macOS and Windows — hold a hotkey, speak, and AI-polished text streams to your cursor. Bring-your-own ASR and LLM, local Qwen3-ASR, and a style-pack marketplace.
StarTrail-org
Apache-2.0 Berkeley SkyLab/BAIR system that renders documents to screenshots and does retrieval entirely in pixel space — up to 18.1% accuracy gain over text RAG with 3x lower token cost.
Mengqi-Lei
An Apache-2.0 generalist model for text-guided object counting across six visual domains, returning instance-grounded points instead of density maps, trained on the 220K-image CLOC benchmark.
localai-org
Apache-2.0 compact neural models for real-time acoustic echo cancellation, noise suppression, and dereverberation on commodity CPUs — 16 ms streaming latency, GGML C++ inference, from 2.7K to 4.8M parameters.
rednote-machine-learning
An Apache-2.0 SGLang extension from RedNote's ML team that classifies attention heads into global/local/retrieval/dense classes to cut long-context prefill FLOPs by 50-70% and speed up TTFT 1.35x-2.2x.
earthtojake
An MIT-licensed library of 11 agent skills that drives real parametric CAD via build123d and OpenCASCADE — STEP output, DXF, G-code slicing, and URDF/SRDF/SDF robot descriptions from plain-language prompts.
openinfer-project
An Apache-2.0 LLM inference engine written purely in Rust and CUDA with no PyTorch — 771 MB resident, 3-second cold start, and flat warm-cache TTFT, serving Qwen3 through trillion-parameter Kimi-K2.
citrolabs
An MIT-licensed macOS browser where you and your AI agents work in parallel — isolated Spaces per agent, inherited Chrome logins, and a JavaScript tool layer any agent CLI can drive.