Project directory

Projects

Open-source AI projects tracked by the leaderboard, ordered by GitHub stars.

Python 175 stars

holarissun/PanelGPT

We introduce new zero-shot prompting magic words that improves the reasoning ability of language models: panel discussion!

### Overview: PanelGPT introduces a new zero-shot prompting method called "Panel Discussion" that enhances language models' reasoning abilities by simulating a panel of experts discussing a topic, improving understanding and perspective. ### Key Features: - **Panel Discussion Prompting**: A method that involves three experts discussing a question step by step, ensuring the result is correct and avoiding penalties. - **Zero-Shot Performance**: Achieves the highest performance on the GSM8K dataset, demonstrating the effectiveness of the Panel Discussion prompt. - **Ablation Studies**: Shows the importance of each element in the Panel Discussion prompt, with ablation studies highlighting the benefits of various components. ### Tech Stack: Python ### Why It Matters: This innovative approach leverages the power of expert knowledge to enhance language models' reasoning capabilities, making it particularly useful in scenarios where consensus is crucial, such as in academic conferences, workshops, and debates.

Python 175 stars

juzhengz/LoRI

[COLM 2025] LoRI: Reducing Cross-Task Interference in Multi-Task Low-Rank Adaptation

Overview: LoRI is a novel variant of LoRA designed specifically for fine-tuning large language models (LLMs). It addresses cross-task interference by freezing the projection matrices and sparsifying the adapter weights, significantly reducing the number of trainable parameters while maintaining single-task performance. Key features: - Freezes projection matrices as random projections - Sparsifies adapter weights using task-specific masks - Reduces the number of trainable parameters - Preserves single-task performance - Minimizes cross-task interference during adapter merging and continual learning Tech stack: Python Why it matters: LoRI simplifies the adaptation process by reducing the computational and memory requirements, making it ideal for fine-tuning large LLMs efficiently and effectively.

Kotlin 175 stars

soniqo/speech-android

On-device speech SDK for Android — ASR, TTS, VAD, and noise cancellation powered by ONNX Runtime with Qualcomm NNAPI acceleration

### Overview: This project provides an on-device speech SDK for Android, powered by ONNX Runtime and Qualcomm NNAPI acceleration. It includes ASR, TTS, VAD, and noise cancellation features, all running locally without cloud APIs or data leaving the device. ### Key Features: - **Streaming Speech Recognition (ASR)**: Supports 25 European languages. - **Text-to-Speech (TTS)**: Offers 8 languages including English, German, and Japanese. - **Voice Activity Detection (VAD)**: Detects voice activity. - **Noise Cancellation**: Reduces background noise. ### Tech Stack: - Kotlin - ONNX Runtime - Qualcomm NNAPI acceleration ### Why It Matters: This SDK is crucial for developers looking to integrate speech capabilities into their Android applications without the need for cloud services or data transfer, ensuring privacy and performance.

Jupyter Notebook 174 stars

Audio-WestlakeU/ATST-SED

This repo includes the official implementations of "Fine-tune the pretrained ATST model for sound event detection".

### Overview: This project includes the official implementations of "Fine-tune the pretrained ATST model for sound event detection," which is accepted at ICASSP 2024. The work is highly related to "ATST" and "ATST-Frame." The project provides a semi-supervised fine-tuning strategy for better utilizing the pretrained model in SED tasks. ### Key Features: - Fine-tune the pretrained ATST model for sound event detection. - Provides a semi-supervised fine-tuning strategy for better using the pretrained model in SED tasks. - Includes inference code for both AudioSet and AudioSetStrong tasks. ### Tech Stack: - Jupyter Notebook ### Why it Matters: This project is highly relevant to the field of sound event detection and fine-tuning of pre-trained models. It provides a practical solution for improving the performance of SED tasks, making it valuable for researchers and practitioners in the field. ### Repository: - Repository: Audio-WestlakeU/ATST-SED - GitHub Stars: 174 - GitHub Description: This repo includes the official implementations of "Fine-tune the pretrained ATST model for sound event detection" (accepted by ICASSP 2024). The work is highly related to "ATST" and "ATST-Frame." ### Updating Notice: - Quick inference script added for easy use. - DESED dataset shared for Chinese users. - Validation dataset definition fixed. - Real dataset download provided. - Strong validation dataset meta files updated. ### Comparing with DCASE Code: - Developed

Python 174 stars

BernieTv/ElevenLabs-Clone

A self-hosted ElevenLabs clone for text-to-speech, voice conversion, and AI audio generation with Docker, FastAPI, and Next.js. 🔊🎙️💡💻

Overview: This project is a self-hosted clone of ElevenLabs, offering text-to-speech, voice conversion, and AI audio generation. It uses Docker, FastAPI, and Next.js, providing a modern, responsive user interface with voice selection, audio history, and user management. Key features: - Fine-tuned TTS with StyleTTS2 - Seamless voice cloning with Seed-VC - Creative audio generation with Make-An-Audio - Fine-tuning for unique voice identities - Dockerized AI stack for easy deployment - FastAPI backend with scalable endpoints - User credit system - Inngest queue to manage workload - AWS S3 integration for audio file storage - Multiple pre-trained voice models - Fully responsive UI with Next.js + Tailwind CSS - Secure authentication with Auth.js - Voice picker component - Audio history tracking Tech stack: Python, Docker, FastAPI, Next.js, AWS S3, Inngest, Tailwind CSS, Auth.js Why it matters: This project is essential for developers looking to create a self-hosted AI audio studio, offering a robust, customizable solution for text-to-speech, voice conversion, and audio generation without relying on external APIs.

Jupyter Notebook 174 stars

FareedKhan-dev/rag-zero-hallucinations

Handling 10M+ docs using RAG with zero hallucinatons

### Overview: Handling 10M+ docs using RAG with zero hallucination is crucial for maintaining the trustworthiness of large-scale document retrieval systems. ### Key Features: - **Zero Hallucination**: The system ensures that answers are generated from the context and verified against the evidence, preventing hallucinations. - **Near-Zero Hallucination**: The model generates answers that are as close to the truth as possible, minimizing the risk of hallucinations. - **Reproducibility**: The process ensures that the results are consistent and can be reproduced, enhancing reliability. ### Tech Stack: - Jupyter Notebook ### Why It Matters: Ensuring that the answers are trustworthy and verifiable is essential for maintaining the integrity and reliability of large-scale document retrieval systems.

Python 174 stars

Sibyl-Labs/Sibyl-Memory

Durable, file-based long-term memory for AI agents. Five-package plugin family: SDK, CLI, MCP server, Hermes adapter, and a LangGraph BaseStore. No vector database, no embeddings.

### Overview: Sibyl Memory is a durable, file-based long-term memory system designed for AI agents. It offers a five-package plugin family including SDK, CLI, MCP server, Hermes adapter, and LangGraph BaseStore. The memory lives on the agent's machine, using a single file on disk, and no vector database or embedding models are utilized. ### Key Features: - **SQLite-backed Five-Tier Hierarchical Schema**: Supports multi-tenant architecture. - **FTS5 Search**: Facilitates efficient FTS5 search capabilities. - **No Vector Database or Embedding Model**: Eliminates the need for vector databases and embedding models. - **No External Retrieval Service**: The memory is self-contained and does not rely on external services. ### Tech Stack: - **Python 3.10+**: The primary language used for development. ### Why It Matters: - **Runs on Linux, macOS (Apple Silicon and Intel), and Windows through WSL2**: Ensures compatibility across different operating systems. - **Privacy Disclosure**: Your memory content never leaves the machine. The plugin talks to `api.sibyllabs.org` for activation, account checks, and size updates, but no personal data is shared. - **Benchmark Performance**: Placed #2 overall in LongMemEval Oracle (ICLR 2025, University of Michigan, 500 questions) with 95.6%, tied with Chronos (PwC), beating MemMachine, Hindsight, Mem0, Supermemory, Zep, and the Oracle baseline.

Python 174 stars

bgts-ai-org/bgts-context-engine

Deterministic code-graph context engine for AI coding agents - PostgreSQL + Apache AGE + pgvector, MCP & REST, no LLM in the loop.

### Overview: BGTS Context Engine is a deterministic code-graph context engine designed for AI coding agents. It leverages PostgreSQL, Apache AGE, pgvector, MCP, and REST to provide precise and reproducible answers to complex questions without relying on large language models. ### Key Features: - **Deterministic Graph Walks:** BGTS Context Engine indexes repositories into a graph, allowing it to answer questions by following the exact paths of symbols, calls, and references. - **Precise Ranking:** The engine ranks answers based on structural information, ensuring that the correct code is returned. - **Context Selector:** Optionally, it provides a context selector that hands the agent the likely-related files in full, making it easier to fix issues in the code. ### Tech Stack: - PostgreSQL - Apache AGE - pgvector - MCP - REST ### Why It Matters: The engine's ability to provide precise and contextually relevant answers to complex questions without relying on large language models makes it highly valuable for AI coding agents. This ensures that the agent can make informed decisions and fix issues in the code with minimal risk of incorrect answers. The deterministic nature of the engine ensures that the same task text, against the same commit, returns the same ranking, providing a consistent and reliable solution. This makes it an ideal tool for developers working on unfamiliar repositories, where the traditional method of embedding search over chunked files fails to provide the necessary structural information.

TypeScript 174 stars

codecentric/c4-genai-suite

[archived] c4 GenAI Suite

Overview: The c4 GenAI Suite is an AI chatbot application that integrates Model Context Provider (MCP) and supports various large language models (LLMs) and multimodal models. It is built with a modern tech stack and designed for extensibility, allowing users to create assistants with different capabilities through extensions. Key features: - Supports OpenAI, Azure OpenAI, Bedrock, Google GenAI, and Ollama compatible models. - Includes a custom RAG server for preparing files and vector stores. - Extensible through MCP servers and extensions like Model Context Protocol (MCP), custom system prompts, Brave Search, and Bing Search. Tech stack: TypeScript Why it matters: The c4 GenAI Suite is designed for developers and administrators to create and manage AI assistants with different capabilities, making it modular and extensible. It supports various models and extensions, ensuring flexibility and customization. Repository: codecentric/c4-genai-suite GitHub stars: 174 GitHub description: [archived] c4 GenAI Suite

Python 174 stars

coleam00/dark-factory-experiment

A repository that ships its own code. AI workflows triage issues, implement them, review, and auto-merge with no human reading the diff. Runs on Archon. The app it maintains is a cited RAG chat over YouTube transcripts, live at chat.dynamous.ai.

### Overview: The Dark Factory Experiment is a public GitHub repository that uses AI coding agents to build, review, and merge code. Humans file issues and promote releases, while AI agents handle the triage, implementation, code review, and testing. ### Key features: - AI-driven issue triage and implementation - Automated code review and merging - Minimal human intervention except for file issues and promote releases - Uses Archon workflow engine for deterministic workflow execution - Runs on Archon, with minimal human interaction ### Tech stack: Python ### Why it matters: This experiment demonstrates the potential of AI-driven development processes, showcasing how AI can automate many tasks traditionally done by humans, reducing the need for human oversight and improving efficiency. The use of Archon workflow engine ensures that the AI agents follow a set of predetermined steps, making the process reliable and traceable.

Rust 174 stars

devwhodevs/engraph

Local knowledge graph for AI agents. Hybrid search + MCP server for Obsidian vaults.

Overview: engraph is a local knowledge graph for AI agents that integrates with Obsidian vaults, providing a 5-lane hybrid search, MCP server, and HTTP REST API. It combines semantic embeddings, full-text search, and LLM-powered reranking for a seamless local experience. Key features: - 5-lane hybrid search - MCP server for AI agents - HTTP REST API - Section-level editing - Vault health diagnostics - Obsidian CLI integration - Real-time sync - Smart write pipeline - Fully local Tech stack: Rust Why it matters: engraph turns Obsidian vaults into a searchable knowledge graph, enabling AI agents to query your notes effectively, understand context, and create new content that fits your vault's structure.

Python 174 stars

inboxpraveen/LLM-Minutes-of-Meeting

Turn audio and video recordings into transcripts, decisions and action items. Supports local models, cloud providers and background jobs you can retry.

### Overview: In today's project, we are developing a comprehensive platform that converts audio and video recordings into structured minutes, decisions, and action items. This tool supports various local and cloud-based models, allowing users to choose their preferred transcription and language processing services. The platform is designed to be highly flexible, running on Docker and offering real-time background jobs for seamless integration and management. ### Key Features: - **Transcription Providers**: Support for Deepgram, AssemblyAI, ElevenLabs Scribe, Sarvam AI, OpenAI-compatible models, and Whisper. - **Language Models**: Compatible with OpenAI, Groq, and local models, including Whisper for local processing without API keys. - **Speaker Labeling**: Automatically attributes decisions and action items to the correct speakers. - **Meeting Length Support**: Transcribes and summarizes long recordings in parts, ensuring no upload limits or model context window constraints. - **Background Jobs**: Celery and Redis with real-time progress tracking, activity logs, and the ability to cancel and retry jobs. - **Configuration**: Configured directly from the app without requiring editing files or restarting the service. - **No Phone Home**: No CDNs, fonts, or telemetry, with outbound calls only to the chosen AI provider. - **Health and Readiness Endpoints**: Built-in health and readiness checks, Prometheus metrics, structured logs, and automatic stale-job recovery. ### Tech Stack: - Python 3.10 or newer - Docker - Flask - Celery - Redis - Whisper (optional for local model processing) - Requirements:

Shell 174 stars

justxor/deepseek-local-guide-ru

🐋 Полный гайд на русском: как установить и запустить DeepSeek локально (V4, R1) на ПК, Mac и сервере — Ollama, llama.cpp, LM Studio, vLLM, Open WebUI, API, RAG, агенты, fine-tuning, лучшие фишки

### Overview: DeepSeek локально — полный гайд и курс на русском (2026): как установить и запустить DeepSeek V4, R1 и дистилляты на ПК, Mac и сервере — бесплатно, приватно и без интернета. ### Key features: - Установка и настройка DeepSeek V4, R1 и дистиллятов - Оllama, llama.cpp, LM Studio, vLLM, Open WebUI, API, RAG, агенты, fine-tuning, лучшие фишки - Полный пошаговый гайд от первой команды до сервера с API, веб-поиском, RAG и агентами ### Tech stack: Shell ### Why it matters: Полный гайд на русском для установки и запуска DeepSeek V4, R1 и дистиллятов на ПК, Mac и сервере, без интернета.

C++ 174 stars

saurabhshri/CCAligner

🔮 Word by word audio subtitle synchronisation tool and API. Developed under GSoC 2017 with CCExtractor.

Overview: A word-by-word audio subtitle synchronisation tool and API designed for developers, focusing on real-time synchronization of subtitles and audio in video files. Key features: - Real-time audio subtitle synchronization - Automatic synchronization of subtitles and audio - Support for various subtitle and audio formats - API for integrating into existing projects Tech stack: C++ Why it matters: This tool is essential for developers working on video editing and subtitle management projects, offering a practical solution for ensuring accurate synchronization between audio and subtitles, thereby improving the overall quality and user experience of their projects.

Python 174 stars

zjunlp/KnowledgeCircuits

[NeurIPS 2024] Knowledge Circuits in Pretrained Transformers

Overview: This project aims to build circuits in pretrained language models responsible for specific knowledge, analyzing their behavior. It includes a demo and a method called EAP-IG for faster computation. Key features: - 🚀 Knowledge Circuits in Pretrained Transformers - 🚀 EAP-IG Method - 🚀 Demo for Knowledge Circuits Tech stack: Python Why it matters: This project provides insights into how large language models acquire new knowledge and analyzes the behavior of specific knowledge circuits, which could be useful for improving model performance and understanding knowledge acquisition mechanisms. Repository: zjunlp/KnowledgeCircuits GitHub stars: 173 GitHub description: NeurIPS 2024 paper on Knowledge Circuits in Pretrained Transformers README excerpt: [Overview], [Installation], [Get the circuit], [Analyze Component], [Acknowledgement], [Citation]

Python 173 stars

FasterDecoding/TEAL

Overview: TEAL is a training-free method for inducing up to 40-50% model-wide activation sparsity in large language models, resulting in significant speedup in single-batch decoding. Key features: - Achieves up to 40-50% model-wide sparsity with minimal performance degradation across Llama-2, Llama-3, and Mistral families. - Compatible with weight quantization, enabling further efficiency gains. Tech stack: Python Why it matters: TEAL simplifies the process of achieving activation sparsity in large language models, making it accessible to a wider range of models and reducing the need for extensive pre-training. Repository: FasterDecoding/TEAL GitHub stars: 173 GitHub description: No description provided. README excerpt: # Training-Free Acivation Sparsity in Large Language Models [[Paper](https://www.arxiv.org/abs/2408.14690)][[Blog](https://www.together.ai/blog/teal-training-free-activation-sparsity-in-large-language-models)] TEAL induces up to 40-50% model-wide activation sparsity in modern LLMs with minimal degradation, resulting in an up to 1.53-1.8x speedup in single-batch decoding. <div align="center"> <img src="figures/clickbait.png" width="500" height="auto"/> </a> </div> The current release supports: - FP16 inference for Llama-

Python 173 stars

HKUDS/XRec

[EMNLP'2024] "XRec: Large Language Models for Explainable Recommendation"

Overview: XRec is a model-agnostic framework that integrates Large Language Models (LLMs) with graph-based collaborative filtering to generate comprehensive explanations for recommendations. Key features: - Integrates graph-based collaborative filtering with LLMs - Establishes a connection between collaborative user-item relationships and language semantics - Supports three public datasets: Amazon-books, Google-reviews, and Yelp Tech stack: Python Why it matters: XRec provides a novel approach to explainable recommendation by leveraging the strengths of both graph-based collaborative filtering and LLMs, offering insights into why certain recommendations are made. Repository: HKUDS/XRec GitHub stars: 173 GitHub description: [EMNLP'2024] "XRec: Large Language Models for Explainable Recommendation"

Python 173 stars

OpenMOSS/MOSS-Video-Preview

An 11B model for real-time video understanding with gated cross-attention and streaming inference

### Overview: MOSS-Video-Preview is an 11B model designed for real-time video understanding, utilizing gated cross-attention and streaming inference. It is built upon the Llama-3.2-Vision architecture, offering state-of-the-art performance in real-time multimodal reasoning. ### Key Features: - 🧩 Image-Video Cross-Attention Architecture: Leverages a native Cross-Attention mechanism to unify image and video understanding, enabling deep decoupling of visual and linguistic features. - 🔄 Low-Latency Interaction & Dynamic Self-Correction: Supports seamless transitions between "Silence" and "Speak" modes, adjusting responses dynamically based on evolving video scenes. - 💡 Extreme Inference Performance & Kernel Optimization: Optimized Cross-Attention kernels and Flash Attention 2 acceleration on CUDA and NPU platforms ensure ultra-low latency with minimal memory overhead. - 📊 Fine-grained Data Synthesis Pipeline: Employs state-of-the-art multimodal LLMs for real-time video understanding, powered by a sophisticated data synthesis pipeline. ### Tech Stack: Python ### Why It Matters: This model is a significant step towards real-time video understanding, offering unparalleled performance and scalability. It is particularly useful in applications requiring immediate analysis and response to video content.

Unknown 173 stars

Rylaispirit/cinematic-video-prompt-skill

AI video prompt cheat sheet & Claude Skill: cinematic camera angles, camera movement, lighting, composition, color grading for Veo 3, Kling, Sora, Runway, Midjourney. 700+ terms with Vietnamese explanations.

Overview: This repository contains a Claude Skill designed to teach AI professionals cinematic vocabulary, including camera angles, movement, lighting, composition, and color grading. It includes 700+ terms in Vietnamese with real-world lessons and examples for various video formats and models. Key features: - 70+ camera angles and shot sizes - 60+ camera movements (dolly, arc, crane, FPV drone, dolly zoom, bullet time) - 80+ lighting terms (golden hour, rim light, Rembrandt, chiaroscuro, volumetric, practical, neon) - 80+ composition rules (rule of thirds, leading lines, S-curve, negative space, repoussoir) - 80+ lens technical film terms (lenses, aperture, shutter, bokeh, flare, filters, film stocks, aspect ratios, color grading tech) - 100+ art styles, 80+ color palettes/grading, 100+ mood/emotion terms - Worked prompt examples for common video types - Character reference sheet prompts for consistency across clips Tech stack: Python, Claude Code, Claude AI, and other AI models that support English prompts. Why it matters: This skill helps AI professionals create more precise and consistent video prompts, improving the quality and coherence of AI-generated videos.

Python 173 stars

Scottcjn/ram-coffers

LLM infrastructure cost reduction via NUMA-aware weight banking: 147 t/s (8.8x stock llama.cpp) on refurbished enterprise POWER8. Self-hosted inference, no cloud APIs. Part of the Proof of Physical AI stack.

### Overview: RAM Coffers is a NUMA-aware conditional memory architecture designed to reduce the cost of Large Language Model (LLM) inference. It routes model knowledge through physical memory banks, enabling self-hosted enterprise hardware like refurbished IBM POWER8 and Apple Silicon-class machines to serve models locally with no cloud API dependency. The system achieves 147 tokens per second (8.8x stock llama.cpp) on POWER8, making it an efficient and zero-abstraction solution for inference. ### Key features: - **NUMA-Distributed Weight Banking**: Model weights are partitioned across NUMA nodes by domain. - **Resonance Routing**: Query embeddings are matched to coffer domain signatures via cosine similarity for intelligent weight activation. - **Non-Bijunctive Pruning**: Selective path collapse before full weight fetch, reducing memory bandwidth requirements. - **DCBT Resident Prefetch**: PowerPC. ### Tech stack: Python. ### Why it matters: RAM Coffers is part of the Proof of Physical AI stack, where real hardware earns real tokens. It achieves 147 tokens per second on POWER8, 8.8x stock llama.cpp, and is verified by the RustChain Proof of Physical AI ecosystem. The system is designed to be self-hosted, eliminating the need for cloud APIs, making it ideal for enterprise environments.