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
mattzh72
An Apache-2.0 agentic system that generates articulated 3D assets as executable Python code rather than meshes, so joints are declared by construction and assets stay editable and forkable.
OpenSenseNova
An Apache-2.0 unified multimodal model series from SenseTime whose NEO-unify architecture drops both the visual encoder and the VAE, with dedicated infographic and interleaved image-text checkpoints.
handy-computer
An MIT-licensed C/C++ speech-to-text engine running 16 ASR model families and 60+ variants on ggml, with GPU backends, GGUF quantization, and per-model WER validation.
Aratako
An MIT-licensed Flow Matching TTS model for Japanese where emoji in the input text steer emotional delivery, with zero-shot voice cloning and caption-based voice design.
stemdeckapp
A free desktop app that splits songs into six stems with Demucs entirely offline, paired with a DAW-style mixer, loop editor, and BPM/key/LUFS analysis.
baidu
Baidu's MIT-licensed OCR model that parses an entire multi-page document in one pass instead of page by page, with vLLM and SGLang serving support.
OpenBMB
OpenBMB's dense 1B on-device LLM claiming 42.57 average across reasoning, code, and agentic benchmarks, shipped with hybrid thinking modes and an open pretraining corpus.
block
Block's self-hostable workspace where humans and AI agents share the same rooms, built on a Nostr relay so every message, patch, and workflow step is one signed event.
fla-org
Triton kernel library implementing 30+ linear, sparse, and state-space attention architectures, verified on NVIDIA, AMD, and Intel — the source of Qwen3-Next's Gated DeltaNet kernels.
XiaomiMiMo
Xiaomi's terminal coding agent, forked from OpenCode, adding SQLite-backed cross-session memory, a deterministic multi-agent workflow runtime, and /dream and /distill self-improvement commands.
img2threejs
An agent skill that rebuilds an object from one reference image as editable, animation-ready Three.js code instead of a mesh file — gated by a staged review pipeline and pure-stdlib Python scripts.
bytedance
ByteDance's 3B active-parameter unified multimodal model handling image and video understanding, generation, and editing in one checkpoint — 85.11 on VBench and 0.90 on GenEval, matching 7B rivals.