AI Trending

Project directory

Projects

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

Python 125 stars

roampal-ai/roampal

Memory that learns what works.

Overview: Roampal is a conversational AI that learns from outcomes, remembers what works, and improves over time, all while maintaining complete privacy and offline functionality. Key features: - Outcome tracking: Scores every result (+0.2 worked, -0.3 failed) - Smart promotion: Good advice becomes permanent, bad advice auto-deletes - Cross-conversation: Recalls from ALL past chats Tech stack: Python Why it matters: Roampal is designed to stop re-explaining yourself every conversation, making interactions more efficient and personalized. It learns from outcomes and improves over time, ensuring that you can do more of what works, while maintaining complete privacy and offline functionality.
Python 125 stars

shootthesound/comfyui-mesh

Split FLUX.2 and LTX 2.3 across two GPUs (LAN or same-machine) — NVENC compresses activations live on the wire. Icarus (ComfyUI node) + Daedalus (back-half server).

### Overview: ComfyUI-Mesh Icarus & Daedalus is a project that splits a diffusion model across two GPUs, leveraging NVIDIA's NVENC to compress activations live. This setup optimizes diffusion models for high-speed networks, reducing wire overhead significantly. ### Key features: - **Split FLUX.2 and LTX 2.3 across two GPUs** using NVENC for live compression. - **Icarus (ComfyUI node) + Daedalus (back-half server)** for efficient diffusion model deployment. - **Supported models** include FLUX.2 Dev, FLUX.2 Klein 9B, and LTX 2.3 (LTX-AV 22B Dev). - **Wire savings** convert directly to wall-clock savings, optimizing diffusion model performance. ### Tech stack: The project leverages Python for its implementation. ### Why it matters: This project optimizes diffusion models for high-speed networks, reducing wire overhead and improving diffusion model performance.
Python 125 stars

silversparro/wav2letter.pytorch

A fully convolution-network for speech-to-text, built on pytorch.

### Overview: A fully convolutional network for speech-to-text, built on PyTorch, implementing the Wav2Letter architecture and utilizing the CTC activation function. ### Key features: - **Train Wav2Letter**: Enables training of the Wav2Letter model. - **Language Model Support**: Integrates language model functionality using KenLM. - **Noise Injection**: Enhances noise robustness through online training. - **Audio Augmentation**: Improves noise robustness through audio augmentation. - **Easy Start/Stop Capabilities**: Provides easy restart or stop functionality during training. - **Visdom/Tensorboard Support**: Visualizes training graphs using Visdom and Tensorboard. - **Training on Raw WAV**: Directly trains the model on raw audio without the need for spectrogram creation. - **Branches for Progress**: Contains code for branches such as `selfAttentionExps` and `trainableFrontEnd` for further development. ### Tech stack: - Python ### Why it matters: This implementation of Wav2Letter provides a robust and efficient solution for speech-to-text tasks, leveraging advanced neural network architectures and robust training techniques to achieve high accuracy and robustness against noise and other challenges.
C# 124 stars

ActiveNick/HoloBot

HoloBot is a reusable 3D interface that allows HoloLens & VR users to interact with any bot using Mixed Reality & Speech.

### Overview: HoloBot is a versatile 3D interface that allows bots to interact with users in Mixed Reality environments, enabling them to be seen, heard, and understood through speech recognition and text-to-speech. ### Key Features: - **Holographic Client**: HoloBot acts as a holographic client for bots, allowing users to interact with them through speech. - **Speech Recognition**: Users can trigger speech recognition by tapping or holding the HoloBot, and the bot responds with a "ping" sound. - **Text-to-Speech**: The bot speaks back to users using speech synthesis, providing a natural language interaction. - **Spatial Sound**: All sounds and speech originate from the HoloBot's location, enhancing immersion. - **Bot Integration**: HoloBot can be integrated with any bot built with the Microsoft Bot Framework, making it a versatile tool for developers. ### Tech Stack: Unity 2018.4.12f1, MRTK v2.1.0, Speech SDK from Microsoft Cognitive Services. ### Why It Matters: HoloBot is a powerful tool that enhances the interaction between bots and users, making them more accessible and engaging through Mixed Reality and speech.
Unknown 124 stars

Azure-Samples/JavaScript-AI-Buildathon

Overview: This repository contains a collection of JavaScript and TypeScript-based AI development quests designed to help developers learn AI through hands-on experience. Key features: - Local AI Development - End-to-End Model Development - Serverless RAG Support System - Build Agents with AI Toolkit - Burger Ordering Agent System Tech stack: JavaScript and TypeScript Why it matters: These quests provide a practical, self-paced learning experience for developers interested in AI development, covering various AI application patterns and agent design. Repository: Azure-Samples/JavaScript-AI-Buildathon GitHub stars: 124 GitHub description: No description provided. README excerpt: ![JavaScript AI Build-a-thon: hands-on Quests, self-paced and always available](assets/javascript-ai-learning-banner.png)
Python 124 stars

Fanziyang-v/FlashVID

[ICLR 2026 Oral] FlashVID: Efficient Video Large Language Models via Training-free Tree-based Spatiotemporal Token Merging

### Overview: FlashVID aims to accelerate video language models (VLLMs) by eliminating the need for training, significantly improving efficiency and performance without compromising accuracy. ### Key Features: - **Training-Free**: FlashVID eliminates the need for training by merging spatiotemporal tokens without compromising model accuracy. - **Efficient Inference**: FlashVID achieves up to **10x** faster inference speed compared to state-of-the-art methods. - **Versatile Compatibility**: Supports inference on multiple VLLMs, including LLaVA-OneVision, LLaVA-Video, and Qwen2.5-VL. ### Tech Stack: Python ### Why It Matters: FlashVID offers a significant improvement in efficiency and performance while maintaining high accuracy, making it a game-changer for VLLMs.
Unknown 124 stars

HKBU-LAGAS/Awesome-Item-ID-Gen-RecSys

Updating curated list of research advancements on item identification and item tokenization in generative recommender systems. The survey is titled "A Survey of Item Identifiers in Generative Recommendation: Construction, Alignment, and Generation"

### Overview: This repository is dedicated to updating and maintaining a curated list of research advancements in item identification and item tokenization within generative recommender systems. The survey titled "A Survey of Item Identifiers in Generative Recommendation: Construction, Alignment, and Generation" is being actively tracked and continuously updated. Contributions to the repository and survey paper are highly encouraged. ### Key features: - **Numeric IDs**: Unified Pretrain, Personalized Prompt & Predict Paradigm (P5) - **Textual IDs**: Elegant Effective Efficient Extensible Solution of Large Language Models for Sequential Recommendation (E4SRec) - **Multi-facet IDs**: How to Index Item IDs for Recommendation Foundation Models (CID) - **Semantic IDs**: Construction, Alignment, and Generation (Survey Paper) ### Tech stack: Python, PyTorch, and Natural Language Processing libraries. ### Why it matters: This repository serves as a valuable resource for researchers, developers, and practitioners interested in the field of generative recommendation systems. It provides comprehensive insights into the current state of research, including the construction, alignment, and generation of item identifiers, which are crucial for improving recommendation systems.
C++ 124 stars

Mobile-Artificial-Intelligence/llama_sdk

lcpp is a dart implementation of llama.cpp used by the mobile artificial intelligence distribution (maid)

Overview: llama_sdk is a Dart implementation of llama.cpp used by the mobile artificial intelligence distribution (maid). Key features: - Optional use of isolations for parallel processing - Streaming support Tech stack: C++ Why it matters: This package provides a Dart implementation for the llama.cpp model, making it easier to integrate llama capabilities into Dart-based projects.
Python 124 stars

OpenMOSS/RoboOmni

A proactive robot manipulation model for multimodal physical environments

### Overview: RoboOmni is a proactive robot manipulation model designed for multimodal physical environments. It leverages end-to-end omni-modal Large Language Models (LLMs) to infer user intentions proactively, enabling effective collaboration between robots and humans. ### Key features: - **Cross-modal Contextual Instructions**: Derives intent from spoken dialogue, environmental sounds, and visual cues. - **Perceiver-Thinker-Talker-Executor Framework**: Unifies intention recognition, interaction confirmation, and action execution. - **OmniAction Dataset**: Contains 140k episodes, 5k+ speakers, 2.4k event sounds, and 640 backgrounds, providing extensive training data for proactive intention recognition. ### Tech stack: - Python ### Why it matters: RoboOmni represents a significant advancement in robotic manipulation by enabling effective collaboration between robots and humans, particularly in real-world interactions where explicit instructions are rare.
Python 124 stars

OraclesTech/guardian-sdk

Ethicore Engine® is an AI safety, ethics, and compliance platform. This repo consists of the open-source components of Ethicore Engine™ - Guardian SDK; designed to protect your AI applications from prompt injection, jailbreaks, role hijacking, system-prompt extraction, and 100+ additional threat categories through a multi-layer analysis pipeline

Overview: Ethicore Engine® Guardian SDK is an open-source component designed to protect AI applications from various threats, including prompt injection, jailbreaks, and role hijacking, ensuring the safety and compliance of your AI models. Key features: - Real-time threat detection for Python LLM and agentic applications - Protection against prompt injection, jailbreaks, adversarial manipulation, and malicious tool calls - Comprehensive threat categories including text, images, audio, and video - Integrations with multiple providers for enhanced security Tech stack: Python Why it matters: Ethicore Engine Guardian SDK is essential for safeguarding AI applications from potential threats, ensuring they operate within ethical and compliant boundaries.
Python 124 stars

ShimShim46/HFT-CNN

Convolutional Neural Network based on Hierarchical Category Structure for Multi-label Short Text Categorization

Overview: This repository contains the implementation of a Convolutional Neural Network (CNN) based on a Hierarchical Category Structure for Multi-label Short Text Categorization. Key features: - Flat model: A non-hierarchical model without fine-tuning. - Without Fine-tuning (WoFt) model: A hierarchical model without fine-tuning. - Hierarchical Fine-Tuning (HFT) model: A hierarchical model with fine-tuning. - XML-CNN model: A Chainer-based implementation of the XML-CNN model by Liu et al. (2017). Tech stack: Python, Chainer, CuPy. Why it matters: This repository provides a comprehensive implementation of a CNN model for multi-label short text categorization, focusing on hierarchical category structure and fine-tuning techniques. It is useful for researchers and developers working on text classification tasks. Repository: ShimShim46/HFT-CNN GitHub stars: 124 GitHub description: Convolutional Neural Network based on Hierarchical Category Structure for Multi-label Short Text Categorization README excerpt: - ![result](https://github.com/ShimShim46/HFT-CNN/blob/media/demo.gif) - These four code/models are Chainer based implementation for text categorization by Convolutional Neural Networks. - * Flat model: Flat non-hierarchical model - * Without Fine-tuning (WoFt) model: Hierarchical model but without Fine-tuning - * Hierarchical Fine-Tuning (HFT) model: Hierarchical and Fine-tuning model
Python 124 stars

aniketmaurya/llm-inference

Large Language Model (LLM) Inference API and Chatbot

### Overview: Aniket Maurya's project provides a Python-based Large Language Model (LLM) inference API and chatbot, powered by LLaMA and Falcon models. The project utilizes the Lit-GPT framework and is designed to be easily integrated into various applications. ### Key Features: - **Inference API**: Offers functionalities for inference of LLaMA and Falcon models. - **Chatbot Interface**: A conversational interface powered by the LitGPT framework, allowing users to interact with the model through a chatbot application. - **Model Preparation**: Users can prepare model weights manually using the `prepare_weights` function. ### Tech Stack: - Python ### Why It Matters: This project simplifies the integration of large language models into applications, making it easier for developers to leverage advanced AI capabilities without the need for extensive model setup and configuration.
HTML 124 stars

gunawan1996/world-forge-ai

AI World Generator 2026: Create Self-Evolving Maps & Stories

Overview: NarrativeDust is an AI-driven narrative engine that transforms a single descriptive sentence into a fully realized, self-animating world. Maps generate from your words, characters emerge with memories and desires, and stories evolve organically, reacting to choices and changes. Key features: - World Genesis from a Single Prompt - Autonomous Character Emergence - Living Storylines (No Fixed Plots) - Dynamic Geography & Resource Systems - Multilingual Narrative Interpretation - Responsive, Device-Agnostic Interface - Adaptive Narrative Intelligence (ANI) - World Branching & Time Travel Tech stack: HTML Why it matters: NarrativeDust turns every idea into an adventure that writes itself, creating immersive worlds that evolve organically based on the choices and actions of the creator.
Python 124 stars

halli75/motion-mirror

Local-first motion transfer — animate any character image from a reference video. Open-source alternative to hosted motion-control tools, built on Wan2.1-VACE.

### Overview: Motion Mirror is an open-source tool designed to animate any character image from a reference video, running entirely on your local machine without cloud services or API keys. It leverages Wan2.1-VACE for high-quality motion transfer, offering whole-body skeleton conditioning for reduced facial and hand flicker. ### Key features: - **Runs Locally**: Supports up to 14B quality on consumer-grade cards, with 9 GB VRAM required for 1.3B backend. - **Whole-Body Motion**: Conditioned on 133 keypoint skeleton, significantly reducing flicker in facial and hand movements. - **Multiple Backends**: Choose between 1.3B, 14B GGUF, and full precision for different trade-offs in speed, identity fidelity, and resource usage. - **CLI, Python API, and Web UI**: Provides a command-line interface, Python API, and a web-based user interface. - **Audio Passthrough**: Passes through the reference video's audio to the output. ### Tech Stack: Motion Mirror is built on Wan2.1-VACE, utilizing its capabilities for high-quality motion transfer. ### Why it Matters: Motion Mirror offers an open-source alternative to hosted motion-control tools, providing a local-first solution that doesn't require cloud services or API keys, making it ideal for developers and content creators looking to streamline their workflow.
Kotlin 124 stars

hiteshsahu/Android-TTS-STT

One line solution for Android Text to speech(TTS) & Speech to Text(STT) translation problem

Overview: This project provides a simple and efficient solution for Android applications to convert text to speech (TTS) and speech to text (STT) without any annoying dialog. Key features: - Text to Speech (TTS) - Speech to Text (STT) - Kotlin-based implementation - Compatible with Android 28 and above - Factory pattern for error handling and callbacks Tech stack: Kotlin Why it matters: This project simplifies the process of integrating TTS and STT functionalities into Android applications, making it easier for developers to implement these features without the need for complex dialogs or permissions. Repository: hiteshsahu/Android-TTS-STT GitHub stars: 124 GitHub description: One line solution for Android Text to speech(TTS) & Speech to Text(STT) translation problem
Python 124 stars

horseee/learning-to-cache

[NeurIPS 2024] Learning-to-Cache: Accelerating Diffusion Transformer via Layer Caching

### Overview: In this project, we introduce **Learning-to-Cache (L2C)**, a novel scheme that enables diffusion transformers to learn to cache layers dynamically. This technique reduces the number of layers required to be kept in memory, significantly improving computational efficiency without compromising model performance. ### Key Features: - **Dynamic Layer Caching**: Optimizes the decision-making process for which layers should be cached to accelerate the diffusion process. - **Significant Performance Improvement**: L2C outperforms existing samplers like DDIM and DPM-Solver across various benchmarks, including DiT-XL/2 and U-ViT-H/2. - **Practical Implementation**: Provides checkpoint files for easy retraining and deployment, with DiT-XL/2 and U-ViT-H/2 examples available. ### Tech Stack: - Python ### Why It Matters: - **Efficiency Gain**: Reduces the computational load by up to 93.68% in cache steps, with minimal impact on FID metrics. - **Performance Enhancement**: Outperforms state-of-the-art samplers like DDIM and DPM-Solver, making it a promising approach for accelerating diffusion transformer models. ### Repository: - **Repository**: horseee/learning-to-cache - **GitHub Stars**: 124 - **GitHub Description**: "NeurIPS 2024" - "Learning-to-Cache: Accelerating Diffusion Transformer via Layer Caching" ### Checkpoint for Routers: - **Models**: DiT-XL/2, U-ViT-H/
Python 124 stars

huangjia2019/rag-project01-framework

[RAG训练营] u.geekbang.org/subject/airag/1009927 这是我用Cursor自行制作的RAG Framework，好处是可以掌控RAG每一个环节的教学细节。

Overview: This project is a self-built RAG Framework, designed to provide a lightweight, customizable knowledge-based question-answering solution. It allows users to construct their own knowledge base and answer questions using their own documents. Key features: - Document Blockization: Supports customizable document block strategies - Vectorization Storage: Converts textual blocks into vectors and efficiently stores them - Similarity-Based Retrieval: Utilizes vector similarity for intelligent matching - Framework Independence: Does not rely on heavy RAG frameworks like LangChain - Cross-Platform Support: Compatible with both Windows and Ubuntu environments Tech stack: Python Why it matters: This project offers a flexible and customizable solution for developing knowledge-based systems, allowing users to tailor the framework to their specific needs without relying on pre-existing RAG frameworks, making it ideal for both educational and professional applications. Repository: huangjia2019/rag-project01-framework GitHub stars: 124
Python 124 stars

jackaduma/LAS_Mandarin_PyTorch

Listen, attend and spell Model and a Chinese Mandarin Pretrained model (中文-普通话 ASR模型)

Overview: This repository contains a PyTorch implementation of the Listen, Attend, Spell (LAS) model, which is a state-of-the-art end-to-end speech recognition (ASR) model. The model is designed to improve upon existing ASR systems by incorporating attention mechanisms to enhance recognition accuracy. Key features: - Implementations of the LAS model - Generation of a vocabulary file for ASR tasks - Training scripts for ASR models - Testing scripts for ASR models Tech stack: PyTorch Why it matters: The LAS model offers significant improvements in speech recognition accuracy by leveraging attention mechanisms, making it a valuable tool for enhancing the performance of ASR systems.
C++ 124 stars

kekzl/imp

Run large AI chat models privately on your own NVIDIA RTX 50 or RTX PRO graphics card. On one RTX 5090, Qwen3.8-27B writes about 70 words per second, 1.4x faster than vLLM. Even a model almost four times larger than the card's memory runs. Works with apps made for ChatGPT or Claude. Starts with one docker command.

### Overview: Imp is a privately hosted AI chat model that runs on NVIDIA RTX 50 or RTX PRO graphics cards. It leverages the power of your own hardware to process large AI chat models, achieving speeds up to 96.3 tokens per second on a 32 GB RTX 5090 card. Imp is designed to be fast and efficient, making it suitable for developers and enthusiasts looking to run large AI models without the need for cloud services. ### Key Features: - **Model Size**: Supports models in the SafeTensors format, including Qwen3.8-27B, Qwen3.6-35B-A3B, and Qwen3.8-Flash-Next. - **Token Speed**: 96.3 tokens/s on a 32 GB RTX 5090 card, which translates to approximately 70 words per second. - **Model Compatibility**: Works seamlessly with apps designed for ChatGPT or Claude, ensuring compatibility with existing chat interfaces. - **One Container**: Imp runs entirely within a single Docker container, eliminating the need for additional installations. - **OpenAI and Anthropic API Integration**: Imp integrates with OpenAI and Anthropic APIs, allowing seamless interaction with popular chat models. - **CUDA 13.4.3**: Utilizes the latest version of the NVIDIA CUDA toolkit for optimal performance. - **MIT License**: The project is released under the MIT license, ensuring freedom and flexibility. ### Tech Stack: Imp is built using C++23 and
Jupyter Notebook 124 stars

mahmoodlab/MMP

Multimodal prototyping for cancer survival prediction - ICML 2024

### Overview: MMP is a multimodal prototyping framework for cancer survival prediction, presented at ICML 2024. It combines gigapixel histology whole-slide images (WSIs) and transcriptomic profiles to predict patient outcomes, offering significant improvements in both accuracy and interpretability compared to existing methods. ### Key Features: - **Morphological Prototypes for WSIs**: Condense gigapixel WSIs into morphological prototypes to reduce tokenization, achieving over 300x compression. - **Biological Pathway Prototypes for Transcriptomics**: Encode transcriptomic profiles with biological pathway prototypes for accurate cellular function characterization. - **Fusion Network**: Process multimodal tokens using a fusion network, either Transformer or optimal transport cross-alignment, operating with a fixed number of tokens for better computation efficiency. ### Tech Stack: - **Jupyter Notebook**: Primary language for development and experimentation. ### Why It Matters: MMP offers a novel approach to survival prediction by leveraging multimodal data, significantly improving prediction accuracy and enabling more interpretable results. The framework's ability to efficiently process large amounts of data without tokenization limitations makes it particularly valuable for complex medical applications. ### Repository: - **GitHub Repository**: mahmoodlab/MMP - **GitHub Stars**: 124 - **GitHub Description**: Multimodal prototyping for cancer survival prediction - ICML 2024 ### Installation: Clone the repository and create a conda environment using the provided `env.yaml` file. ### MMP Walkthrough: 1. **Dataset Organization**: Organize dataset files