AI Trending

Project directory

Projects

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

C++ 732 stars

ddiakopoulos/tinyply

:earth_africa: lightweight C++17 ply 3d mesh format importer & exporter

### Overview: tinyply is a lightweight, public domain implementation of the PLY mesh file format. It is designed for easy import and export of PLY files in both binary and ASCII formats, supporting files sizes up to 4GB and handling big-endian binary files. The library is written in C++17 and requires a recent compiler. ### Key Features: - **Binary and ASCII Support**: Supports importing and exporting PLY files in both binary and ASCII formats. - **High Performance**: Fast parsing for common import schemes, optimized for performance. - **Public Domain**: Released under the Unlicense, making it freely usable and distributable. - **Compatibility**: Works with files sizes up to 4GB and can handle big-endian binary files. - **Performance Trade-off**: Simplified design for ease of use, potentially at the cost of performance. ### Tech Stack: - **Language**: C++ - **Dependencies**: Minimal, requires C++17 and a recent compiler. ### Why it Matters: tinyply is a valuable tool for computer vision, graphics, and data visualization applications that require efficient and flexible handling of PLY files. Its simplicity and performance make it a preferred choice for many projects, especially in scenarios where performance is a secondary concern. ### Repository: - **Repository**: ddiakopoulos/tinyply - **Stars**: 732 - **Description**: A single-header, zero-dependency implementation of the PLY mesh file format, suitable for various applications in the computer vision and graphics communities. ### Getting Started: The project includes
Shell 732 stars

shannhk/hermes-agent-control-room

Control Room-first template for managing Hermes agents from one VPS agent to specialist teams and orchestrated workflows

Overview: This is a public template designed to manage Hermes agents, providing a control plane for agents, orchestrators, and specialists, allowing for a smooth transition from a single agent to a specialized team or an orchestrated workflow. Key features: - Agent Control Room structure - Agent runbooks and Docker notes - Secret maps and backups - Level-based architecture for growth - Task bus pattern for delegation - Bundled setup and operations skills Tech stack: Shell Why it matters: This template simplifies the management of Hermes agents, making it easier to scale from a single agent to a team or an orchestrated workflow, ensuring a clean and organized system. Repository: shannhk/hermes-agent-control-room GitHub stars: 732 GitHub description: Control Room-first template for managing Hermes agents from one VPS agent to specialist teams and orchestrated workflows.
PowerShell 731 stars

JulietMirambo/Units_of_Measure_Harmonization-intelligence-platform

Production-Grade ML System for Automated Unit of Measure Error Detection | 88-92% Accuracy | 94% Autonomy | KNIME Workflow

### Overview: Manufacturing and procurement organizations lose millions annually due to Unit of Measure (UOM) errors. A single misplaced decimal or wrong unit can lead to order fulfillment disasters, inventory chaos, supply chain disruptions, financial losses, and compliance issues. This project introduces a production-grade Machine Learning System for Automated UOM Error Detection, offering 88-92% accuracy and 94% autonomy, built into a KNIME workflow. ### Key Features: - **Real-Time Error Detection**: Identifies UOM issues as they occur. - **Confidence Scoring**: Provides ML probability scores for error classification. - **Auto-Correction Tracking**: Tracks the system's progress in fixing errors. - **Root Cause Analytics**: Understands the reasons behind UOM errors. - **Intuitive Visualizations**: Color-coded, charts, statistics, and performance metrics. - **Performance Metrics**: Speed, accuracy, and autonomy. - **Built into KNIME Workflow**: Zero additional setup needed. ### Why It Matters: Manufacturing and procurement organizations are losing millions annually due to UOM errors. This project offers a solution to prevent these losses by detecting, classifying, and correcting UOM errors in real-time, providing a 94% autonomy rate and 88-92% accuracy. The system is integrated into a KNIME workflow, making it easy to implement without additional setup. ### Tech Stack: - **ML Engine**: 60+ engineered features, XGBoost classifier, 5-fold cross-validation. - **Processing**: 3,300 records per
Python 731 stars

mit-han-lab/distrifuser

[CVPR 2024 Highlight] DistriFusion: Distributed Parallel Inference for High-Resolution Diffusion Models

### Overview DistriFusion is a training-free algorithm designed to accelerate high-resolution diffusion model inference on multiple GPUs without compromising image quality. It addresses the fragmentation issue of naive patch-based approaches by employing synchronous communication for patch interaction at the initial step. Subsequently, it leverages asynchronous communication to reuse activations from the previous step, reducing communication overhead. ### Key Features - **Synchronous Communication for Initial Patch Interaction**: Ensures seamless patch interaction from the start. - **Asynchronous Communication for Subsequent Steps**: Utilizes previous activations for faster computation. - **High-Resolution Inference**: Achieves 1.8x, 3.4x, and 6.1x speedups for 50-step DDIM sampler at 1280×1920 resolution on 2, 4, and 8 A100 GPUs, respectively. ### Tech Stack - **Python 3** - **NVIDIA GPU + CUDA >= 12.0** ### Why It Matters - **Speedup**: Significantly reduces inference latency while maintaining high-resolution image quality. - **Quality Preservation**: Ensures visual fidelity even with increased resolution. - **Training-Free**: No need for additional training, making it ideal for production environments. - **Versatility**: Works with various diffusion models and can be integrated into existing pipelines. ### GitHub Repository - **Repository**: [GitHub Repository](https://github.com/mit-han-lab/distrifuser) - **Stars**: 731 - **Description**: [CVPR 20
Python 731 stars

staruhub/ClaudeSkills

13 curated Agent Skills for research, product decisions, decks, publishing, audits, and more — portable across skills-compatible agents.

Overview: ClaudeSkills is a collection of Agent Skills designed to streamline research, product decisions, and more, providing a portable solution across compatible agents. Key features: - Deep Research: Scopes the question, gathers sources, and checks citations. - Product Manager: Turns vague ideas into a product document. - Deck Studio: Generates an 8-slide consulting deck. - WeChat Article Writer: Writes WeChat articles with copy, image prompts, and layout HTML. Tech stack: Python Why it matters: ClaudeSkills provides a structured approach to research, product development, and content creation, making it a valuable tool for researchers, product managers, and content creators. Repository: staruhub/ClaudeSkills GitHub stars: 732 GitHub description: 13 curated Agent Skills for research, product decisions, decks, publishing, audits, and more — portable across skills-compatible agents.
Python 730 stars

NVlabs/Long-RL

Long-RL: Scaling RL to Long Sequences (NeurIPS 2025)

### Overview: Long-RL: Scaling RL to Long Sequences is a project that introduces a framework designed to scale up reasoning in vision-language models (VLMs) to long videos. The project leverages reinforcement learning to address the unique challenges of long video reasoning, integrating a large-scale dataset, a two-stage training pipeline, and a specialized training infrastructure. ### Key Features: - **Large-Scale Dataset:** LongVideo-Reason, comprising 104K long video QA pairs with high-quality reasoning annotations across diverse domains. - **Two-Stage Training Pipeline:** Extends VLMs with chain-of-thought supervised fine-tuning (CoT-SFT) and reinforcement learning (RL). - **Training Infrastructure:** Multi-modal Reinforcement Sequence Parallelism (MR-SP), incorporating sequence parallelism and a vLLM-based engine tailored for long video, using cached video embeddings for efficient rollout and prefilling. ### Tech Stack: - Python ### Why It Matters: - **Performance Improvement:** LongVILA-R1-7B achieves strong performance on video benchmarks, reaching 65.1% and 71.1% accuracy on VideoMME without and with subtitles, respectively, and consistently outperforming LongVILA-7B across multiple benchmarks. - **Efficiency:** Supports processing up to 8,192 video frames per video, and configurable FPS settings. - **Speedup:** Achieves up to 2.1x speedup on long video RL training. - **Public Availability:** Releases a training system for public availability that supports
Python 730 stars

lucidrains/TimeSformer-pytorch

Implementation of TimeSformer from Facebook AI, a pure attention-based solution for video classification

Overview: TimeSformer is a pure attention-based solution for video classification, achieving state-of-the-art performance. This implementation is based on the research from Facebook AI, focusing on the 'Divided Space-Time Attention' variant. Key features: - Pure attention-based approach - Focuses on time axis attention before spatial attention - Includes a diagram for visual understanding Tech stack: Python Why it matters: This implementation showcases a simple yet effective solution for video classification, demonstrating the potential of pure attention-based models. Repository: lucidrains/TimeSformer-pytorch GitHub stars: 730 GitHub description: Implementation of TimeSformer from Facebook AI, a pure attention-based solution for video classification. README excerpt: <img src="./diagram.png" width="600px"></img>
TypeScript 729 stars

Maneek21/Deft

Self-hostable AGPL workspace where humans and AI agents share chat, tasks, knowledge, approvals, and MCP context.

Overview: Deft is a self-hostable AGPL workspace where humans and AI agents can collaborate on chat, tasks, knowledge, approvals, and more. Key features: - Collaborative chat - Task proposal and approval - Sharing of knowledge and calendar context - Integration with AI agents Tech stack: TypeScript Why it matters: Deft provides a platform for humans and AI agents to work together efficiently, leveraging the strengths of both to enhance productivity and decision-making. Repository: Maneek21/Deft GitHub stars: 725 GitHub description: Self-hostable AGPL workspace where humans and AI agents share chat, tasks, knowledge, approvals, and MCP context.
Unknown 729 stars

gabriben/awesome-generative-information-retrieval

### Overview: This repository aims to provide a comprehensive overview of conversational models and generative information retrieval, focusing on two main topics: **Grounded Answer Generation** and **Generative Document Retrieval**. It includes various techniques and tools for enhancing the capabilities of chatbots and information retrieval systems, aiming to make them more effective and reliable. ### Key features: - **Grounded Answer Generation (RAG)**: Techniques for improving the accuracy of generated answers by incorporating external sources. - **LLM Memory Manipulation**: Grounding the model's internal weights to enhance the quality of generated content. - **Re-Ranking**: Methods for improving the relevance of search results. - **Self-Correction**: Techniques to correct errors in generated content. - **Fact Uncertainty Estimates**: Assessing the reliability of information retrieved. - **Constrained Generation**: Ensuring the generated content aligns with specific constraints. - **Data Centric**: Methods that prioritize the quality of data used. - **Utility Maximization**: Techniques that optimize the overall utility of the generated content. - **Multimodal**: Incorporating multiple types of data for better retrieval. - **Prompting**: Techniques for generating specific prompts. - **Generate Code**: Generating code snippets as answers. - **Query Generation**: Methods for generating queries from user inputs. - **Summarization and Document Rewriting**: Techniques for generating concise summaries and rewriting documents. - **Table QA**: Answering questions based on tables. - **Continual Learning for Generative Retrieval**: Techniques for improving the system's ability to learn from new data
TypeScript 729 stars

swarmclawai/swarmvault

The local-first LLM Wiki: open-source knowledge graph builder, RAG knowledge base, and agent memory store. Built on Andrej Karpathy's pattern. An Obsidian alternative for personal knowledge management, AI second brain, and durable Claude Code / Codex / OpenClaw memory.

### Overview: SwarmVault is a local-first knowledge graph builder, RAG knowledge base, and agent memory store. It turns your documents, code, transcripts, notes, and URLs into a durable markdown wiki and local graph, serving as an AI second brain and durable Claude Code/Codex/OpenClaw memory. ### Key Features: - **Local-first Knowledge Graph**: Converts your sources into a persistent knowledge graph. - **RAG Knowledge Base**: Uses a rich-ankle graph to enhance search capabilities. - **Markdown Wiki**: Generates markdown pages and outputs for easy access. - **Context Packs**: Provides context-rich summaries for quick reference. - **Graph Viewer**: A local graph viewer for inspecting and querying the knowledge graph. - **CLI Interface**: Command-line interface for quick and easy setup and management. ### Tech Stack: TypeScript ### Why It Matters: SwarmVault turns your documents into a durable and searchable knowledge graph, making it easier to manage and access your knowledge. It's an alternative to Obsidian for personal knowledge management, research, and AI second brain functionality. ### Repository: Repository: swarmclawai/swarmvault GitHub Stars: 730 ### Languages: - English - Chinese - Japanese ### Try It in 30 Seconds: ```bash npm install -g @swarmvaultai/cli swarmvault quickstart ./your-repo ``` ### Documentation: Documentation on the website is currently English-first. If wording drifts between translations, `README.md` is the canonical source. ### Three-Layer Architecture
TypeScript 729 stars

tech-systems/panes

🎉📱 Create dynamic modals, cards, panes for your applications in few steps. Any framework and free.

### Overview: Cupertino Panes is a lightweight, multi-functional modals, cards, and panes library that simplifies the creation of dynamic UI elements for modern applications. It is designed to be easily integrated into any framework and is free of dependencies, making it perfect for small projects and large-scale applications alike. ### Key Features: - **Small Size:** 12kb (minified and gzipped bundle with all modules). No dependencies. - **Reusable:** Create your configuration once and use it everywhere. - **Modularized:** Add extra features to your panes and create your own modules. - **Accelerated:** Hardware-accelerated transitions and native behavior. - **Progressive:** Useful for mobile, web, and hybrid applications. ### Tech Stack: TypeScript ### Why It Matters: - Cupertino Panes is a crucial tool for developers looking to enhance the user experience of their applications with modern, touch-friendly UI elements. - It simplifies the process of creating dynamic modals, cards, and panes, making it easier to integrate into existing projects. - The library is lightweight, making it suitable for both small and large-scale applications, ensuring that developers can focus on the core functionality of their applications without being bogged down by unnecessary dependencies. ⭐ We appreciate your star, it helps!
JavaScript 728 stars

Aryanutkarsh/SeeCode

Animated, explorable diagrams for Claude and other coding agents, from a short JSON spec.

Overview: SeeCode is an interactive diagramming tool that converts a short JSON specification into an animated, explorable diagram. It supports various diagram types and can export the diagram in multiple formats. Key features: - Animated diagrams - Explorable and exportable - Supports various diagram types - Can be exported as SVG, PNG, GIF, or MP4 Tech stack: JavaScript Why it matters: SeeCode simplifies the process of creating and visualizing diagrams, making it easier for developers to communicate complex systems and processes visually.
Python 728 stars

ltjed/freephdlabor

freephdlabor: customizing personalized multiagent systems that researchs 24/7 on your own scientific problem

Overview: freephdlabor is an open-source multiagent framework designed to automate the entire scientific research lifecycle—from hypothesis generation through experimentation to publication-ready manuscripts. It provides a dynamic, customizable, and human-in-the-loop approach to research, enabling researchers to focus on their ideas and insights. Key features: - Dynamic Workflows: Adapts in real-time to research findings. - Fully Customizable: Easily add, modify, or remove agents via built-in customization support. - Human-in-the-Loop: Naturally integrates human feedback in real-time. - Continual Research: Context management enables sustained exploration. Tech stack: Python. Why it matters: freephdlabor offers a streamlined, automated solution for scientific research, making it accessible and efficient for researchers across various fields. Repository: ltjed/freephdlabor GitHub stars: 728 GitHub description: freephdlabor: customizing personalized multiagent systems that research 24/7 on your own scientific problem GitHub link: https://github.com/ltjed/freephdlabor
Python 728 stars

salesforce/xgen

Salesforce open-source LLMs with 8k sequence length.

### Overview: XGen is an official research release of Salesforce AI Research, featuring a 7B language model trained on an 8K input sequence length. This model is designed for advanced research purposes and supports both 4K and 8K sequence lengths, enabling it to handle longer input sequences effectively. ### Key features: - **Model Cards**: Available on HuggingFace Hub for easy access and integration. - **Model Versions**: - `XGen-7B-4K-Base`: Supports 4K sequence length. - `XGen-7B-8K-Base`: Supports 8K sequence length. - `XGen-7B-8k-Inst`: Instruction-finetuned for research purposes. ### Tech stack: The models are built using the OpenAI Tiktoken package, which can be installed via `pip`. ### Why it matters: This release is crucial for researchers and developers looking to advance the field of language modeling and AI research. The ability to train models on longer sequences and the availability of multiple model versions make XGen a valuable tool for academic and industrial applications.
Unknown 728 stars

wagamamaz/tensorflow-tutorial

TensorFlow and Deep Learning Tutorials

### Overview: This repository contains a collection of TensorFlow tutorials and resources for deep learning and machine learning. It serves as a comprehensive guide for developers looking to understand and implement various deep learning models and techniques. ### Key Features: - TensorFlow Official Deep Learning Tutorial - Introduction to TensorFlow and its applications - Various deep learning models such as MLP with Dropout, Autoencoder, CNN, RNN, Deep Reinforcement Learning, Sequence to Sequence, Word Embedding, and more. - Resources for deep learning reading list and tutorials ### Tech Stack: - TensorFlow ### Why It Matters: This repository is essential for developers looking to dive into deep learning and TensorFlow. It provides a structured approach to learning and implementing various deep learning models, making it a valuable resource for anyone interested in the field.
Python 728 stars

yeyupiaoling/MASR

Pytorch实现的流式与非流式的自动语音识别框架，同时兼容在线和离线识别，目前支持Conformer、Squeezeformer、DeepSpeech2模型，支持多种数据增强方法。

Overview: MASR是一款基于Pytorch实现的自动语音识别框架，支持流式和非流式识别，兼容在线和离线识别，目前支持Conformer、Squeezeformer、DeepSpeech2模型，以及多种数据增强和预处理方法。 Key features: - 支持多个语音识别模型 - 支持多种解码器 - 支持多种预处理方法 - 支持多种数据增强方法 - 支持多种推理方法 Tech stack: Python Why it matters: MASR简化了语音识别项目的部署和使用，适用于服务器、Nvidia Jetson设备以及未来计划支持的Android等移动设备。
TypeScript 727 stars

0xK3vin/MegaMemory

Persistent project knowledge graph for coding agents. MCP server with semantic search, in-process embeddings, and web explorer.

### Overview: MegaMemory is a persistent project knowledge graph designed for coding agents, providing a semantic search and in-process embeddings for efficient querying and updating of concepts. ### Key Features: - Persistent project knowledge graph - Semantic search over the knowledge graph - In-process embeddings for efficient querying and updating - Supports natural language queries and exact ID lookup - Built with TypeScript and Node.js - npm package available: `megamemory` - GitHub stars: 726 - GitHub description: MCP server with semantic search, in-process embeddings, and web explorer ### Tech Stack: TypeScript, Node.js ### Why it Matters: MegaMemory helps coding agents remember across sessions by building and querying a graph of concepts, architecture, and decisions directly from the code. This approach eliminates the need for static analysis and AST parsing, making it more efficient and accurate.
Python 727 stars

ARahim3/mlx-dspark

Up to 4× faster LLM decoding on Apple Silicon, lossless. Native MLX port of DeepSeek's DSpark & z-lab's DFlash speculative decoding — Gemma-4, Qwen3.8, Muse-Glimmer, Nemotron, LFM2.5, Ornith-1.0, ternary Bonsai-27B.

### Overview: MLX-DSpark is a powerful tool designed to accelerate the decoding process of large language models (LLMs) on Apple Silicon, offering up to 4x faster performance compared to traditional methods. This project leverages the native support for DeepSeek's DSpark and z-lab's DFlash speculative decoding on Apple Silicon, ensuring lossless output and compatibility with various models. ### Key Features: - **Native Support for DSpark and DFlash**: MLX-DSpark runs both DeepSeek's DSpark and z-lab's DFlash speculative decoding on Apple Silicon, providing significant performance improvements. - **Lossless Output**: The drafters ensure that the output is identical to normal decoding, making it suitable for lossless applications. - **Versatile Models**: Supports a wide range of models, including Gemma-4, Qwen3, LFM2.5, Muse-Glimmer, Ornith-1.0, and others, with measured speedups ranging from 2.5x to 4.31x. ### Tech Stack: - Python ### Why It Matters: MLX-DSpark is crucial for developers and model enthusiasts looking to optimize the performance of their LLMs on Apple Silicon. By leveraging the native support for DSpark and DFlash, MLX-DSpark can significantly reduce decoding times, making it an essential tool for accelerating LLM usage on Apple devices. ### Repository: Repository: ARahim3/mlx-dspark GitHub Stars: 719 Documentation: [arahim.dev/mlx-dspark](https://
Python 727 stars

kyegomez/LongNet

Implementation of plug in and play Attention from "LongNet: Scaling Transformers to 1,000,000,000 Tokens"

### Overview: LongNet is an open-source implementation of the paper "LongNet: Scaling Transformers to 1,000,000,000 Tokens" by Jiayu Ding, Shuming Ma, Li Dong, Xingxing Zhang, Shaohan Huang, Wenhui Wang, and Furu Wei. It is a Transformer variant designed to scale sequence length up to 1 billion tokens without sacrificing performance on shorter sequences. ### Key Features: - **Dilated Attention Mechanism**: Utilizes a dilated attention mechanism to handle longer sequences efficiently. - **Efficient Memory Usage**: Designed to scale up to 1 billion tokens while maintaining performance on shorter sequences. - **Versatile Applications**: Suitable for a wide range of applications, including natural language processing, computer vision, and more. ### Tech Stack: - **Python**: The primary programming language used for implementation. - **PyTorch**: The deep learning framework utilized for building and training the model. ### Why It Matters: LongNet is significant because it addresses the scalability challenge of transformers by enabling them to process extremely long sequences. This makes it a valuable tool for applications where the amount of data is enormous, such as in large language models, image recognition systems, and more. By improving the performance of transformers on longer sequences, LongNet opens up new possibilities for innovation and efficiency in various fields. ### Repository: The project is hosted on GitHub at [kyegomez/LongNet](https://github.com/kyegomez/LongNet). It has been starred by 727 users and is
TypeScript 727 stars

mmpneo/curses

Speech to Text and KB input captions for OBS, VRChat, Twitch chat and Discord

### Overview: Curses is a powerful tool for creating captions and captions for OBS, VRChat, Twitch chat, and Discord. It offers a wide range of customization options, including speech-to-text and text-to-speech capabilities, as well as support for various social media platforms. ### Key Features: - **Speech to Text and Text to Speech**: Supports Microsoft Azure, Speechly, Deepgram, and WebSpeechApi for speech-to-text, and Microsoft Azure, Uberduck, and Windows API for text-to-speech. - **OBS Captions Customization**: Offers customization options for colors, fonts, shadows, background textures, text typing animation, sound effects, and particle effects. - **Google Fonts**: Utilizes over 1000 free fonts for OBS captions. - **OBS Stream Customization**: Allows for the customization of OBS stream captions. - **VRChat**: Supports KillFrenzy Avatar text and native captions. - **Twitch**: Enables the use of 7TV/FFZ/BTTV emotes, post STT to chat, and use chat messages as captions and TTS. - **Discord**: Sends STT to a specified channel. - **Scenes**: Saves multiple designs and automatically switches designs when OBS changes scene. ### Tech Stack: TypeScript ### Why it Matters: Curses is essential for anyone looking to enhance their captions and captions for OBS, VRChat, Twitch chat, and Discord. Its comprehensive features and customization options make it a valuable tool for streamers, game developers, and anyone involved in live content creation.