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
QwenLM
Alibaba's latest Qwen LLM, focused on stability and real-world utility with stronger agentic, repository-level coding and persistent reasoning across turns.
nvidia
NVIDIA's Speech-Augmented Language Model that tops the Open ASR Leaderboard at 5.63% WER, pairing a FastConformer encoder with a Qwen LLM decoder.
OpenBMB
OpenBMB's tokenizer-free TTS system, now a 2B model trained on 2M+ hours across 30 languages with Voice Design, controllable cloning, and 48kHz audio.
facebookresearch
Meta's MIT-licensed deep-learning library for audio processing and generation. Bundles the state-of-the-art EnCodec neural audio codec with MusicGen, a controllable music generation model supporting both textual and melodic conditioning — a foundational toolkit for open-source audio AI.
tencent-ailab
Tencent AI Lab's open-source song generation system (LeVo) built on a hybrid LM-plus-diffusion architecture and multi-preference alignment. Generates full songs up to 4m30s from structured lyrics and style tags, with dual-track vocal/accompaniment modeling and multilingual support.
ace-step
MIT-licensed open-source music generation foundation model that pairs a Qwen3-based LM planner with a Diffusion Transformer decoder. Runs text-to-music, cover generation, repainting, and track separation locally across NVIDIA, AMD, Intel, and Apple hardware, generating a full song in under two seconds on an A100.
roboflow
Apache-2.0 real-time transformer detector from Roboflow, built on a DINOv2 backbone and presented at ICLR 2026. Delivers SOTA accuracy/latency trade-offs on COCO and RF100-VL, handles object detection and instance segmentation in one API, is designed for fine-tuning, and installs with a single pip command — a license-friendly alternative to AGPL YOLO models.
janhq
Apache-2.0 desktop app from Menlo Research that runs LLMs like Llama, Gemma, Qwen, and GPT-oss fully offline as a privacy-first ChatGPT alternative — with optional cloud routing to OpenAI, Anthropic, Mistral, and Groq, custom assistants, MCP agentic support, and an OpenAI-compatible local server at localhost:1337.
datalab-to
GPL-3.0 open-source engine from Datalab that converts PDF, image, PPTX, DOCX, XLSX, HTML, and EPUB documents into clean Markdown, JSON, chunks, and HTML — with structure-aware parsing of tables, math, and forms, optional LLM-assisted accuracy, JSON-schema structured extraction, and ~25 pages/sec batch throughput on an H100.
Blaizzy
MIT-licensed package for running and fine-tuning Vision Language Models on Apple Silicon using MLX — supports Qwen2-VL, Gemma 4, Phi-4, MiniCPM, DeepSeek-OCR, Pixtral, LLaVA and more, with CLI, Python, Gradio, and FastAPI interfaces, speculative decoding (DFlash, EAGLE-3, MTP) for 2-4x speedup, KV cache quantization, LoRA/QLoRA fine-tuning, and distributed inference across multiple Macs.
BerriAI
MIT-licensed Python SDK and self-hosted AI Gateway that exposes 100+ LLM providers — OpenAI, Anthropic, Gemini, Bedrock, Azure, VertexAI, vLLM, NVIDIA NIM, and more — through an OpenAI-compatible interface, with virtual keys, spend tracking, load balancing, fallbacks, guardrails, and observability callbacks for Lunary, MLflow, and Langfuse.
vllm-project
Apache-2.0 omni-modality inference engine from the vllm-project organization — extends vLLM's PagedAttention and continuous batching to text, image, video, audio, TTS, and diffusion workloads with full disaggregation, heterogeneous pipelines, and CUDA/ROCm/MUSA/NPU/XPU backends.