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
Anionex
A CLI toolkit and agent skill that gives text-only coding agents like DeepSeek working vision — image Q&A, pixel-coordinate grounding, long-screenshot OCR and UI restoration — by putting the multimodal capability in the harness rather than the model.
LiteReality
An agentic pipeline that turns a single iPhone RoomPlan scan into an interactable 3D room of articulated objects, using a deterministic scene-init stage followed by a Claude Code or Codex agent that edits the scene until it matches the capture.
microsoft
Microsoft's official ggml runtime for VibeVoice-ASR-BitNet, using ternary I2_S quantization on the decoder and I8_S on the VAE to shrink a 4.62 GB ASR model to 1.58 GB and run it faster than real time on ordinary CPUs.
prathoshap
A Sanskrit chant TTS system from IISc Bengaluru that sings ślokas with metrically-aware durations rather than reading them flat — F5-TTS DiT backbone, BigVGAN-v2 vocoder, and a reusable Sanskrit text frontend with sandhi and meter detection.
MiniMax-AI
MiniMax's open-weights text-to-music model generating full five-minute songs at 32 kHz stereo, using an 8B global LLM for structure, a 0.6B local LLM for acoustic detail, and Flow Matching synthesis from fused hidden states.
jd-opensource
JD's real-time streaming video editor: a 16B multimodal diffusion transformer that edits live video frames causally as they arrive from a text instruction, hitting 30 FPS at 720x1248 in its deployment benchmark, with Apache-2.0 weights on Hugging Face.
MakazhanAlpamys
A one-YAML fine-tuning CLI whose layer streaming trains Llama-3.1-8B on a 4 GB laptop GPU at a measured 3.32 GB peak, plus a 'soup ship' release gate that scores a tune against seven offline suites and returns SHIP or DON'T SHIP as an exit code.
NVIDIA-NeMo
NVIDIA's pre-alpha Rust proxy that translates between OpenAI Chat, Anthropic Messages, and OpenAI Responses so Claude Code or Codex can run against vLLM, NIM, or Ollama — with classifier, stage, escalation, and random routing plus Prometheus metrics.
google-deepmind
DeepMind's open weights and JAX code for WeatherNext 2 and WeatherNext Cyclones — 0-15 day forecasts in under a minute on one TPU, three-day cyclone forecasts as accurate as prior models' two-day, and a 1° Mini model that runs on a free Colab TPU runtime.
lightningpixel
A local-first desktop app that turns images or prompts into 3D meshes entirely on your own GPU, shipping no model itself — generators like Hunyuan3D 2 Mini, TripoSG, and Trellis2 GGUF install as swappable GitHub extensions.
QwenLM
Alibaba's official QwenLM plugin suite that makes text-only agent harnesses multimodal-native — eight capabilities, each installed as a Skill plus optional MCP server, across Claude Code, Codex, Gemini CLI and more.
semantica-agi
A deterministic, LLM-free graph layer that sits under your agent stack turning every fact and decision into an auditable node — W3C PROV-O provenance, Rete/Datalog/SPARQL reasoning, and polyglot RDF plus property-graph storage.