AI Trending

Project directory

Projects

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

Jupyter Notebook 103 stars

ukairia777/pytorch-nlp-tutorial

pytorch를 사용하여 텍스트 전처리부터 RAG, 에이전트, LLM 파인튜닝을 정리한 Deep Learning NLP 저장소입니다.

Overview: 이 저장소는 PyTorch를 활용하여 텍스트 전처리, RAG, 에이전트, LLM 파인튜닝 등 다양한 NLP 작업을 정리한 딥 러닝 NLP 교과서입니다. Key features: - PyTorch를 사용한 텍스트 전처리 - RAG (Recurrent Auto-encoder Generation) 모델 구현 - 에이전트 (Agent) 학습 - LLM (Large Language Model) 파인튜닝 Tech stack: PyTorch Why it matters: 이 저장소는 딥 러닝 기술을 활용한 NLP 작업의 기초적인 이해와 실습을 제공하며, 실제 프로젝트에서 활용할 수 있는 기술을 정리한 콘텐츠입니다.
Go 102 stars

GaeaRuiW/kube-llmops

Overview: kube-llmops is a comprehensive Kubernetes-native LLM operations platform that provides a unified interface for deploying, managing, monitoring, and optimizing your LLM infrastructure. Key features: - Model Serving: Supports vLLM, llama.cpp, and TEI engines. - CLI (`kubectl-llmops`): Provides 15+ commands for deploying, listing, scaling, and more. - AI Gateway: Unified OpenAI-compatible API with key management and rate limiting. - Observability: Prometheus, Grafana, Langfuse, and node-exporter. - Logging: Fluent Bit, Loki, and Grafana Explore. - Autoscaling: KEDA with scale-to-zero fallback. - Security: Keycloak SSO, LLM-Guard, and NetworkPolicy isolation. - RAG Infrastructure: Dify platform, pgvector, TEI embedding/reranking, Ragas evaluation, and quality gate. - Fine-tuning: LLaMA-Factory LoRA/QLoRA/Full fine-tuning with Argo Workflows. - Model Distribution: MinIO model cache, HuggingFace fallback, and hf-transfer multi-threaded downloads. - Storage: MinIO S3-compatible model storage and PVC model cache. Tech stack: Go, Helm, Kubernetes, Prometheus, Grafana, Langfuse, Fluent Bit, Loki, KEDA, Keycloak, NetworkPolicy, Dify, pgvector, TEI embedding/reranking, Ragas evaluation, quality gate, Argo Workflows, MinIO, HuggingFace, hf-transfer, MinIO S3-compatible model
Python 102 stars

INV-WZQ/ReactiveGWM

[Arxiv 2026] ReactiveGWM: Flexible Control and NPC Reactivity in Game World Models

### Overview: ReactiveGWM is a flexible control and NPC reactivity model designed for game world models. It assigns control roles to individual characters at rollout initialization, using spatial role binding and unified agency conditioning with causal self-attention to enhance character interactions. ### Key Features: - **Spatial Role Binding**: Grounds learned handles in initial-frame instance masks. - **Unified Agency Conditioning**: Binds each character's external action sequence or conditional NPC rule to the same handle. - **Causal Self-Attention**: Allows characters to respond to the current and preceding video context. ### Tech Stack: - Python ### Why it Matters: ReactiveGWM provides a flexible and effective way to control characters in game models, enhancing their interactions and reactivity. This model is particularly useful in scenarios requiring dynamic and adaptive character behavior, such as in game simulations or virtual environments. ### GitHub Stars: 102 ### Repository: - Repository: INV-WZQ/ReactiveGWM - Primary Language: Python - GitHub Stars: 102 ### Updates: - **[Sep 29, 2026]**: v2 HNM and SF3 training and inference code released, with Hugging Face model/data support, custom character controls, and role reassignment. The original release remains available on the [v1 branch](https://github.com/INV-WZQ/ReactiveGWM/tree/v1). ### Introduction: ReactiveGWM combines character-specific spatial binding and agency conditioning with causal video context to enhance character interactions in game models. This model is
JavaScript 102 stars

Stormycry-cryp/dsh-AuthInOne

Self-contained DeepSeek Harness (DSH) plugin for Provider/Auth login, model switching, image fallback, token/cost analytics, and same-port Web restart. Useful? A star helps.

### Overview: dsh-AuthInOne is a self-contained DeepSeek Harness plugin designed for seamless login, model switching, and token analytics within DSH. It supports OpenAI Codex, Kimi Code, and other compatible providers, offering optional vision fallback and detailed analytics. ### Key Features: - **Provider Login**: Seamless integration for various providers. - **Model Switching**: Easy model switching without replacing DSH's per-session selector. - **Vision Fallback**: Optional fallback for text-only models. - **Token Analytics**: Detailed token usage and cost tracking. - **Cost Analytics**: Real-time cost tracking within DSH's native settings. ### Tech Stack: JavaScript ### Why It Matters: A star on GitHub helps support the project, ensuring the plugin remains useful and maintained for developers. ### Repository: GitHub: [Stormycry-cryp/dsh-AuthInOne](https://github.com/Stormycry-cryp/dsh-AuthInOne) ### GitHub Stars: 102 ### README Excerpt: dsh-AuthInOne is a DeepSeek Harness plugin for Provider login, API and custom OpenAI-compatible Provider setup, model switching, optional vision fallback for text-only models, token usage analytics, and cost tracking inside DSH's native **Models** and **Usage** settings. OpenAI Codex uses browser OAuth with state, S256 PKCE, and a loopback callback. Kimi Code opens the Provider's complete authorization link backed by an RFC 8628 device flow; the link carries the short code, so
Python 102 stars

To3akaRin/VLMForge

Evaluate visual models on your own images, JSON Schema, and production constraints with LangGraph, Pareto analysis, and a no-key Replay demo.

Overview: VLMForge is a decision workbench designed for teams building structured extraction with cloud vision models, evaluating each provider route, structured-output strategy, parameter set, latency, and cost against the same frozen prompt, schema, preprocessing, and private image contract. Key features: - Evaluates visual models on images, JSON Schema, and production constraints. - Uses LangGraph for orchestration, Pareto analysis, and no-key replay demo. - Provides schema and critical-rule failures, never hidden by an LLM judge or an attractive aggregate score. - Freezes real production contract into reproducible runs. - Includes a no-key Replay adapter for demos, CI, and bug reports. Tech stack: Python. Why it matters: VLMForge helps teams discover models that can extract private catalog attributes, honor cross-field rules, meet P95 targets, and stay within per-item budgets, providing a more comprehensive evaluation than public benchmarks. Repository: To3akaRin/VLMForge GitHub stars: 102 GitHub description: Evaluate visual models on your own images, JSON Schema, and production constraints with LangGraph, Pareto analysis, and a no-key Replay demo.
Jupyter Notebook 102 stars

WASasquatch/easydiffusion

Easy Diffusion is an advanced Stable Diffusion Notebook with a feature rich image processing suite.

Overview: Easy Diffusion is an advanced Stable Diffusion Notebook designed for easy image processing and text-to-image generation. Key features: - Text to Image Stable Diffusion - Image to Image Stable Diffusion - Inpainting Stable Diffusion - NSFW Filtering (Gaussian Blurred Images) - Cached Pipes for faster loading - Optional Attention Slicing - HF Model Support - Stable Diffusion Concept Support - Use local or remote images/masks/prompt files - Noodle Soup Prompts support - Random word support - Negative prompt support - Recursive Evolution - Image Upscaling (GOBIG, IMG2IMG, GFPGAN, CodeFormer, Real-SRGAN) - Image Processing (Sharpening, Chromatic Aberration, Median Filter, Depth Output, Fake DOF, Tileable Texture Output) - CLIP Interrogate Tech stack: Jupyter Notebook Why it matters: Easy Diffusion simplifies the process of generating high-quality images from text descriptions, making it accessible for both beginners and advanced users. Its features make it a versatile tool for various image processing tasks, from basic to advanced, and it supports a wide range of models and techniques, making it a valuable asset in the field of image generation and processing.
Jupyter Notebook 102 stars

alvarobartt/serving-pytorch-models

Serving PyTorch models with TorchServe :fire:

Overview: This project serves PyTorch models using TorchServe, a framework developed by PyTorch for model serving. It demonstrates the process of training and deploying a transfer learning CNN model using ResNet as the backbone, specifically classifying images from a Food101 dataset slice. Key features: - Training and deploying a transfer learning CNN model - Using ResNet as the backbone - Serving the model with TorchServe - Utilizing PyTorch's implementation of ResNet Tech stack: Python, PyTorch, TorchServe, Java Why it matters: This project showcases how to efficiently deploy and serve machine learning models using a production-grade framework, demonstrating the practical application of PyTorch and TorchServe in real-world scenarios.
Python 102 stars

awslabs/rhubarb

A Python framework for multi-modal document understanding with Amazon Bedrock

Overview: Rhubarb is a Python framework designed for building document and video understanding applications using Multi-modal Large Language Models and Embedding models. It leverages Amazon Bedrock and supports foundation models such as Anthropic Claude and Amazon Nova for document processing and Amazon Titan for embeddings. Key features: - Document Q&A - Streaming chat with documents (Q&A) - Document Summarization (Page, Full, Specific Pages, Streaming) - Structured data extraction - Extraction Schema creation assistance - Named entity recognition (NER with 50 built-in entities) - PII recognition - Figure and image understanding from documents - Large document processing with sliding window approach - Document Classification with vector sampling using multi-modal embedding models - Logs token usage to track costs Tech stack: Python Why it matters: Rhubarb simplifies the process of integrating Amazon Bedrock models into Python applications, making it easier to build applications that can understand and process documents and videos using multi-modal models.
Rust 102 stars

cyberlife-coder/VelesDB

The explainable, local-first memory engine for AI agents. One ~9 MB binary fuses vector + graph + columnar under VelesQL; why() returns the evidence path behind every recall. No cloud, no glue code — runs on server, browser, mobile and desktop.

Overview: VelesDB is a local-first memory engine designed for AI agents, combining vector, graph, and columnar data structures under a single query language. It is optimized for performance and cost efficiency, running on servers, browsers, mobile, and desktop without the need for cloud services or glue code. Key features: - Local-first memory engine - Runs on server, browser, mobile, and desktop - Supports vector, graph, and columnar data structures - Provides evidence path for every recall - No cloud, no glue code, no LLM, no API key in memory path Tech stack: Rust Why it matters: VelesDB offers a compact, efficient solution for AI agents that requires a persistent memory without the overhead of cloud services or glue code. It simplifies the development process by eliminating the need for complex integration with cloud-based solutions, making it ideal for developers and researchers looking to build AI applications with minimal infrastructure requirements.
Rust 102 stars

elevenlabs/cli

Command Line Interface to interact with ElevenLabs voice agents

Overview: The ElevenLabs CLI is a command-line interface designed for interacting with ElevenLabs voice agents, providing full API access and agent management capabilities. Key features: - Full API access to all ElevenLabs endpoints - Agents as Code, allowing management from local configuration files - Supports Conversational AI agents - Data residency ensures agent configurations are kept locally Tech stack: Rust Why it matters: The CLI simplifies interaction with the ElevenLabs platform, making it easier to manage and utilize voice agents programmatically. Repository: elevenlabs/cli GitHub stars: 102
Python 102 stars

kamwoh/partcraft

[ECCV2024] PartCraft: Crafting Creative Objects by Parts

### Overview: PartCraft is a generative visual AI model that allows users to craft creative objects by selecting visual concepts in parts. This approach differs from traditional methods by enabling users to choose objects by parts, resulting in fine-grained and faithful generation. ### Key Features: - **Part Discovery**: PartCraft parses objects into parts through unsupervised feature clustering. - **Textual Representation**: Parts are encoded into text tokens using an entropy-based normalized attention loss. - **Bottleneck Encoder**: A bottleneck encoder projects part tokens, enhancing fidelity and accelerating learning. ### Tech Stack: - **Python** - **PyTorch** - **Hugging Face** - **DreamCreature** ### Why It Matters: PartCraft revolutionizes creative control in generative visual AI by enabling users to select visual concepts by parts, leading to highly customized and innovative creations. The model's ability to generate holistically faithful and plausible results sets it apart from traditional methods.
Kotlin 102 stars

lu0973-cb02d/ClassCue

按课表记好每一节课、被点名时及时提醒的本地优先 Android 课堂助手：离线录音转写、点名提醒、课后回放搜索与可选 AI 答案

### Overview: 课醒 ClassCue 是一款专为 Android 设备设计的课堂助手应用，旨在帮助学生按课表记录每一节课，并在被点名或提问时及时提醒。它提供了离线录音转写、点名提醒、课后回放搜索以及可选的 AI 答案功能，旨在提高学习效率和课堂参与度。 ### Key features: - **按课表记录课程**：自动根据星期、上下课时间、起止周和单双周录入课程，可填写地点。 - **离线录音转写**：同一路麦克风音频同时写入本地 WAV 和离线语音识别器。 - **点名与提问提醒**：综合学号、姓名、称呼、点名说法和提问说法打分判断；支持同音或近音字识别。 - **课后回放搜索**：跨课程搜索老师原话、提醒和答案。 - **导出功能**：单节课、一门课或全部课程导出为 Markdown。 - **省空间与防丢失**：课后录音压缩为 AAC 格式，保留原件，应用关闭或重启时自动修复未完成的录音。 ### Tech stack: Kotlin ### Why it matters: 课醒 ClassCue 是一款功能强大且实用的课堂助手应用，帮助学生有效管理课程、及时提醒，并提供丰富的学习资源。通过离线录音转写、点名提醒和课后回放搜索等功能，提高了学习效率和课堂参与度。
C++ 102 stars

microsoft/MSVBASE

MSVBASE is a system that efficiently supports complex queries of both approximate similarity search and relational operators. It integrates high-dimensional vector indices into PostgreSQL, a relational database to facilitate complex approximate similarity queries.

Overview: MSVBASE is a system designed to efficiently support complex queries involving approximate similarity search and relational operators on both scalar and vector data-sets, integrating high-dimensional vector indices into PostgreSQL. Key features: - Supports approximate similarity search and relational queries - Integrates high-dimensional vector indices into PostgreSQL - Compatible with PostgreSQL syntax and protocol - Compatible with hnsw and sptag indices - Supports spann and more indices in the future - Compatible with PostgreSQL SQL syntax Tech stack: C++ Why it matters: MSVBASE enhances the capabilities of PostgreSQL by integrating advanced vector similarity search and relational query processing, making it a valuable tool for applications requiring complex data analysis and retrieval.
TypeScript 102 stars

stabgan/openrouter-mcp-multimodal

MCP server for OpenRouter — chat with 300+ LLMs (Claude, Gemini, GPT), analyze images / audio / video, generate images / speech / music / video (Veo 3.1, Seedance, Kling, Wan) from Claude Desktop, Cursor, Kiro, VS Code.

### Overview: OpenRouter MCP Multimodal is a production-grade Model Context Protocol (MCP) server that connects AI coding agents to OpenRouter's unified LLM API over stdio. It includes 21 tools and supports 300+ OpenRouter models for text, vision, audio, and video analysis and generation. ### Key Features: - **21 Tools:** Supports various AI tools including Claude Desktop, Cursor, GPT, Gemini, GPT, Kiro, VS Code, and more. - **300+ Models:** Covers 300+ OpenRouter models for comprehensive AI capabilities. - **Unified API:** Connects to OpenRouter's unified LLM API over stdio, providing a seamless experience. ### Tech Stack: - TypeScript ### Why It Matters: OpenRouter MCP Multimodal is essential for developers looking to integrate AI tools into their projects, offering a comprehensive solution for text, vision, audio, and video analysis and generation. Its inclusion in the MCP Registry ensures its reliability and compatibility with other MCP servers. Repository: stabgan/openrouter-mcp-multimodal Primary Language: TypeScript GitHub Stars: 102
JavaScript 102 stars

unstonio/pixelgpt-24x24

Tiny local text-to-24x24 pixel art model, trained on roughly 200K samples in 30 minutes on an RTX 5090.

Overview: PixelGPT 24x24 is a tiny, local text-to-pixel art model trained on 200K samples, generating 24x24 sprites with a five-color palette in 30 minutes on an RTX 5090. It's designed to be fully self-contained, using only local inference without sending data to an API. Key features: - Local inference with no external data sent - 24x24 canvas and five-color palette - Trained on 200K samples in 30 minutes - 68M-parameter autoregressive pixel transformer - Tight format for pixel art generation Tech stack: JavaScript Why it matters: PixelGPT 24x24 demonstrates the potential of text-to-pixel art generation, showcasing how a small model can produce high-quality pixel art with a visible generation process. This model is designed to be fully local, making it ideal for small devices and environments where internet access is limited or unavailable.
Python 101 stars

Freim32/vllmops

Self-hosted vLLM made simple. Git-friendly YAML for models and profiles, full lifecycle from an intuitive CLI and TUI.

Overview: vLLMops simplifies the setup and management of vLLM servers, providing a Git-friendly YAML configuration and a live TUI for easy control. Key features: - Git-friendly YAML for models and profiles - Full lifecycle management via CLI and TUI - Single gateway for all models - Live metrics and logs for each model - Per-project virtual environment Tech stack: Python Why it matters: Simplifies the deployment and management of vLLM servers, making it easier to set up and control multiple models in a project. Repository: Freim32/vllmops
Python 101 stars

G-U-N/UniRL

[ICML 2026] a unified reinforcement learning toolbox for joint RL on language models and diffusion models

### Overview **PromptRL** is a unified reinforcement learning toolbox designed for joint reinforcement learning on language models and diffusion models. It addresses the limitations of current flow-based RL pipelines by incorporating language models as adaptive prompt refiners. This release introduces two RL paths: **image editing** with **FLUX.1-Kontext-dev** and **FLUX.2-klein-base-4B**, and **text-to-image GenEval** with **FLUX.1-dev** and **FLUX.2-klein-base-4B**. Each path offers a prompt-only trainer and a joint LLM+DiT trainer. ### Key Features - **Image Editing**: Uses **FLUX.1-Kontext-dev** and **FLUX.2-klein-base-4B** for training and evaluation. - **Text-to-Image GenEval**: Uses **FLUX.1-dev** and **FLUX.2-klein-base-4B** for training and evaluation. - **Prompt-Only Trainer**: Optimizes the prompt refiner with frozen vision and text encoders. - **Joint Trainer**: Combines language model and diffusion model training for improved performance. ### Tech Stack - **Python 3.11** - **PyTorch 2.7** - **CUDA 12.x** - **NVIDIA GPUs** ### Why It Matters PromptRL addresses the limitations of current flow-based RL pipelines by integrating language models as prompt refiners. This approach helps overcome exploration collapse and prompt overfitting, leading to more diverse and effective image generation
Python 101 stars

ML-GSAI/Scaling-Diffusion-Transformers-muP

[NeurIPS 2025] Official implementation for our paper "Scaling Diffusion Transformers Efficiently via μP".

### Overview: This repository contains the official implementation of our NeurIPS 2025 paper "Scaling Diffusion Transformers Efficiently via μP". The paper explores the use of μP, a method for stabilizing hyperparameter (HP) transfer across different model widths, to improve the efficiency and performance of diffusion Transformers. ### Key Features: - **μP (Maximal Update Parametrization)**: A method for stabilizing optimal hyperparameters across different model widths, enabling direct transfer of HPs from small models to large models. - **DiT-XL-2-μP**: A variant of DiT-2 that uses μP for HP stabilization, resulting in improved performance and reduced tuning costs. - **PixArt-α-μP**: A variant of PixArt-α that uses μP, demonstrating robust HP transferability and faster convergence. ### Tech Stack: Python ### Why It Matters: The use of μP in diffusion Transformers leads to significant improvements in performance and reduced tuning costs, making it a promising solution for large model pretraining. The repository provides the necessary code and setup to replicate the experiments on DiT and PixArt-α, following the original DiT and PixArt-α codebases and the Maximal Update Parametrization (μP) methodology. ### Systematic Investigation for DiT-μP on ImageNet: - **Implementation of DiT-μP**: Based on the original codebase, we adjust the weight multiplier and variance of each weight according to μP theory, using the MuAdamW optimizer from the original
Python 101 stars

ShiZhengyan/DePT

[ICLR 2024] This is the repository for the paper titled "DePT: Decomposed Prompt Tuning for Parameter-Efficient Fine-tuning"

Overview: This repository contains the code for the paper titled "DePT: Decomposed Prompt Tuning for Parameter-Efficient Fine-tuning," which presents a method to improve parameter efficiency in fine-tuning language models without significantly increasing the number of trainable parameters. Key features: - Decomposed Prompt Tuning (DePT) - Supports Llama-2 models - Integrates seamlessly with parameter-efficient transfer learning Tech stack: Python Why it matters: DePT offers a way to enhance the performance of parameter-efficient fine-tuning without compromising the model's performance, making it a valuable tool for researchers and developers working with large language models.
Python 101 stars

batmanlab/Mammo-CLIP

[MICCAI 2024, top 11%] Official Pytorch implementation of Mammo-CLIP: A Vision Language Foundation Model to Enhance Data Efficiency and Robustness in Mammography

### Overview: Mammo-CLIP is an official PyTorch implementation of Mammo-CLIP, a vision-language foundation model designed to enhance data efficiency and robustness in mammography. ### Key Features: - **Enhanced Data Efficiency**: Mammo-CLIP leverages a large dataset to improve model performance. - **Robustness in Mammography**: The model is specifically tailored for enhancing data efficiency and robustness in mammography tasks. - **Official Implementation**: A PyTorch implementation of Mammo-CLIP, providing a reliable and reproducible solution. ### Tech Stack: Python ### Why it Matters: Mammo-CLIP is a significant advancement in the field of mammography, offering improved data efficiency and robustness in image classification tasks. Repository: batmanlab/Mammo-CLIP GitHub Stars: 101