Project directory

Projects

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

Python 25 stars

Liyan06/ChartMuseum

[NeurIPS 2025] ChartMuseum: Testing Visual Reasoning Capabilities of Large Vision-Language Models

### Overview: ChartMuseum is a chart question answering benchmark designed to evaluate reasoning capabilities of large vision-language models (LVLMs) over real-world chart images. The benchmark consists of 1162 image, question, and short answer tuples, focusing on non-trivial textual and visual reasoning skills. It is collectively annotated by a team of 13 researchers in computer science. ### Key features: - **Benchmark Design**: ChartMuseum is a comprehensive chart question answering benchmark. - **Annotation Team**: Co-created by a team of 13 researchers in computer science. - **Image Quality**: Consists of 1162 image, question, and short answer tuples. - **Reasoning Skills**: Focuses on non-trivial textual and visual reasoning skills. ### Tech stack: Python ### Why it matters: ChartMuseum serves as a valuable tool for evaluating the reasoning capabilities of large vision-language models, providing insights into how these models perform in real-world scenarios involving charts.

Rust 25 stars

MaincodeHQ/mainarch

A from-scratch Rust inference stack for AMD GPUs that talks straight to the Linux kernel driver. No ROCm, HIP, or HSA runtime. Serves OLMo 2 on MI355X.

Overview: mainarch is a Rust-based inference stack designed for AMD GPUs, bypassing ROCm, HIP, and HSA runtimes. It directly interacts with the Linux kernel driver, serving Open Language Modeling (OLMo) 2 on a MI355X. Key features: - Direct interaction with the Linux kernel driver - Skip layers like ROCm, HIP, and HSA runtime - Open `/dev/kfd` and build the dispatch packet by hand - Serve OpenAI-shaped demo page at `http://127.0.0.1:8080` Tech stack: Rust Why it matters: mainarch provides a reference architecture and prototype for serving large language models on AMD GPUs without vendor runtimes, offering a direct and efficient approach to inference on the hardware. Repository: MaincodeHQ/mainarch GitHub stars: 25 GitHub description: A from-scratch Rust inference stack for AMD GPUs that talks straight to the Linux kernel driver. No ROCm, HIP, or HSA runtime. Serves OLMo 2 on MI355X.

Python 25 stars

MontyGovernance/montycat-mcp

Shared, persistent memory for AI agents. Self-hosted MCP server with semantic search, vector RAG, and live updates. Works with Claude, Cursor, Codex, and any MCP client.

### Overview: Montycat MCP is a self-hosted Model Context Protocol (MCP) server that provides persistent, searchable memory for AI agents. It allows Claude, Codex, and Cursor to share and access shared memories across multiple sessions. ### Key Features: - **Memory Persistence**: Decisions, preferences, and project context are carried over chat sessions. - **Shared Memory**: All MCP clients use the same memory, ensuring consistency. - **Searchable Memory**: Supports vector search, keyword recall, and hybrid recall modes. - **Real-Time Updates**: Live updates ensure that all agents see the most current information. ### Tech Stack: Python ### Why it Matters: Montycat MCP offers a robust, self-hosted solution for AI agents, eliminating the need for hosted memory services or cloud embedding APIs. It ensures data integrity and privacy by running on the user’s machine, making it ideal for teams that want to control their own memory and data. ### Repository: MontyGovernance/montycat-mcp - **GitHub Stars**: 26 - **GitHub Description**: A shared, persistent memory for AI agents. Self-hosted MCP server with semantic search, vector RAG, and live updates. Works with Claude, Cursor, Codex, and any MCP client. - **Primary Language**: Python - **GitHub Description**: Shared, persistent memory for AI agents. Self-hosted MCP server with semantic search, vector RAG, and live updates. Works with Claude, Cursor, Codex, and any MCP client.

Python 25 stars

Sinapsis-AI/sinapsis-chatbots

Monorepo for sinapsis templates supporting LLM based Agents

Overview: Sinapsis Chatbots is a monorepo for Sinapsis chatbot packages, integrations, and demo webapps, designed to support LLM-based agents. Key features: - Sinapsis Anthropic: Templates for text-to-text, image-to-text, and tool-enabled chatbots using Anthropic's Claude models. - Sinapsis Chatbots Base: Shared chatbot foundation for building and processing LLMConversationPacket flows. - Sinapsis ChatHistory: Integration for retrieving and managing chat history. - Sinapsis LLaMA-CPP: Template for LLM chat workflows using the LLaMA-CPP model. - Sinapsis LLaMA-Index: Integration for LLaMA-Index for enhanced chatbot capabilities. - Sinapsis Mem0: Template for chat processing with Mem0 model. - Sinapsis VLLM: Template for VLLM-based chat workflows. Tech stack: Python Why it matters: Sinapsis Chatbots simplifies the development and deployment of LLM-based chatbots, providing a unified workspace for shared chatbot abstractions, integrations, and demo webapps, making it easier to build, test, and deploy chatbot solutions.

Python 25 stars

SkyuForever/CRE-LLM

CRE-LLM: A Domain-Specific Chinese Relation Extraction Framework with Fine-tuned Large Language Model

Overview: CRE-LLM is a domain-specific Chinese relation extraction framework that leverages a fine-tuned large language model for efficient and accurate relation extraction tasks. Key features: - Fine-tuned large language model for relation extraction - Support for FinRE and SanWen benchmarks - Data processing and preparation for both FinRE and SanWen datasets - Supervised fine-tuning of language models for triple generation Tech stack: Python Why it matters: CRE-LLM provides a powerful solution for Chinese relation extraction tasks, offering efficient and accurate results through the use of a fine-tuned large language model.

HTML 25 stars

TitanByDark/resp-codex-image-context-runtime

A Codex plugin and local MCP runtime for context-bounded image generation and inspection.

Overview: A Codex plugin and local MCP runtime for context-bounded image generation and inspection. Key features: - Context-bounded image generation - Image inspection - Integration with Codex Tech stack: HTML Why it matters: This plugin and runtime provide a powerful tool for generating and inspecting images within a specific context, enhancing the capabilities of Codex in generating high-quality images for various applications.

Python 25 stars

YuanYeYouTao/Yuki

An open-source, self-hosted social AI agent across QQ, Telegram, Feishu/Lark and personal clients, with persistent identity, long-term memory, autonomous participation, a Linux workspace and resumable multi-agent tasks.

### Overview: Yuki 是一个开源、自托管的社会化 AI Agent，通过 QQ、Telegram、飞书、Lark 和个人客户端参与对话，记住人物、偏好和共同经历，也能在持久 Linux 环境里处理文件、运行程序、维护项目和执行定时任务。人物身份、记忆与任务由自己的数据库保存，更换模型、账号或网关后可以继续使用。 ### Key features: - **多平台接入**：QQ、Telegram、飞书、Lark 和个人客户端共用身份、记忆、工具和任务。 - **自主行动**：支持长任务续跑和多平台协作。 - **持久记忆与任务管理**：人物身份、记忆与任务由自己的数据库保存，更换模型、账号或网关后可以继续使用。 - **代码代理**：独立[本机连接器](docs/operations/local-agent-connector.zh-CN.md)让服务器上的 Yuki 查询和监督本人授权的 Codex、Claude Code 或 ACP Agent，进度和问题交回原任务。 - **多平台接入**：飞书和 Lark 使用企业自建应用的长连接，支持单聊、群聊和话题，接入方式见[配置说明](docs/operations/feishu.zh-CN.md)。 - **多平台接入**：个人设备可以配对、发送附件、接收事件、查看任务，并向已有任务补充信息；桌面项目 [Yuki Desktop](https://github.com/Institute-of-Agent-Life/Yuki-Desktop)

Jupyter Notebook 25 stars

awslabs/guidance-for-multi-omics-and-multi-modal-data-integration-and-analysis-on-aws

This guidance creates a scalable environment in AWS to prepare genomic, clinical, mutation, expression and imaging data for large-scale analysis and perform interactive queries against a data lake. The solution also demonstrates the use of Amazon Omics for multi-modal analysis.

### Overview: This guidance provides a scalable environment on AWS to prepare genomic, clinical, mutation, expression, and imaging data for large-scale analysis. It also demonstrates the use of Amazon Omics for multi-modal analysis. ### Key Features: - **Genomic Data Conversion**: Libraries for converting genomics data. - **Multi-Modal Data Preparation**: Serverless data ingestion pipelines for multi-modal data preparation. - **Interactive Queries**: Visualize and explore clinical data through an interactive interface. - **AWS Omics Integration**: Demonstrates how to use AWS Omics for multi-modal analysis. ### Tech Stack: Jupyter Notebook ### Why it Matters: This solution simplifies the integration and analysis of complex multi-modal data, making it accessible for large-scale genomic and clinical studies.

Python 25 stars

botAGI/AGmind-ML

Locally fine-tuned Russian RAG models: document splitter, query expansion, retrieval embedder. Teacher distillation → GGUF → llama.cpp on AMD Vulkan. Commercial-OK licenses.

### Overview: AGmind-ML is a locally fine-tuned Russian RAG models project, including a document splitter, query expansion, and retrieval embedder. The models are trained on a single consumer GPU and deployed on AMD Strix Halo, using GGUF inference and commercial-OK licenses. ### Key features: - **Document Splitter**: Context-aware document splitter for RAG, preserving semantic chunks, tables, and code. - **Query Expansion**: Search-query expansion for qmd, improving English hallucinations on Russian. - **Retrieval Embedder**: RU-first no-prefix dense retriever, 4 layers, distilled by layer-pruning + BGE-M3 hard negatives. ### Tech stack: Python ### Why it matters: AGmind-ML provides locally fine-tuned Russian RAG models for the self-hosted AGmind RAG stack, with teacher distillation, quantization to GGUF, and inference on AMD Vulkan via llama.cpp. The models are trained on a single consumer GPU and deployed on AMD Strix Halo, making them fully local and commercial-OK. Repository: botAGI/AGmind-ML

Python 25 stars

catalina17/VideoNavQA

An alternative EQA paradigm and informative benchmark + models (BMVC 2019, ViGIL 2019 spotlight)

Overview: VideoNavQA is an innovative visual question answering (VQA) task designed to bridge the gap between visual and embodied question answering by removing navigation and action selection requirements. This task focuses on increasing the difficulty of visual reasoning through a larger question space, challenging the existing methods of evaluating EQA feasibility. Key features: - A novel VQA task that removes navigation and action selection requirements from embodied QA. - The dataset is available for download [here](https://drive.google.com/drive/folders/1DpEdjmVDMeJZ0ohS_TTp0HAjEbX0fU_m?usp=sharing). - Supports evaluation of models using FiLM-based approaches. Tech stack: Python Why it matters: VideoNavQA provides a unique framework for evaluating the feasibility of visual question answering tasks, highlighting the complexities of visual reasoning in a challenging setting. Repository: catalina17/VideoNavQA Primary language: Python GitHub stars: 25 GitHub description: An alternative EQA paradigm and informative benchmark + models (BMVC 2019, ViGIL 2019 spotlight) ![Dataset statistics](https://github.com/catalina17/VideoNavQA/blob/master/samples/dataset%20stats.png)

Rust 25 stars

chenhunghan/mlx-training-rs

A CLI in Rust to generate synthetic data for MLX friendly training

### Overview: A CLI tool for generating synthetic data suitable for MLX-friendly training. The CLI is based on the PHP version and is designed for fine-tuning models on Apple Silicon. ### Key features: - Generates synthetic data for MLX training - Compatible with MLX fine-tuning - Supports 4-bit quantized MLX model - Requires Python 3.11 and HomeBrew installation ### Tech stack: - Rust - MLX - Python 3.11 - HomeBrew ### Why it matters: This tool simplifies the process of fine-tuning models on Apple Silicon, making it easier for developers to experiment with local LLM fine-tuning without the need for complex setup or expensive hardware.

Elixir 25 stars

elchemista/vettore

Vettore is a small vector toolkit for Elixir that keeps your data in ETS and uses Rust only where it helps. It's small, fast, efficient, simple!

### Overview: Vettore is a small vector toolkit for Elixir that leverages ETS for efficient storage and Rust for acceleration, particularly in specialized operations like SIMD/GPU distance kernels and HNSW search. It provides a simple and efficient way to integrate Elixir with Rust for performance-critical tasks. ### Key Features: - **ETS-backed collections**: Data is stored in ETS, making it easy to inspect and manage. - **Exact flat search**: Provides exact matches for vectors. - **Native HNSW approximate search**: Supports approximate nearest neighbors. - **Matryoshka-style funnel search**: Facilitates hierarchical search. - **Binary quantized candidate search**: Efficient search for binary vectors. - **Hybrid candidate pipelines**: Combines exact and approximate search results. - **ColBERT-style late interaction**: Handles multi-vector records. - **MUVERA-style fixed-dimensional encodings**: Efficient encoding for vectors. - **Named distance, similarity, normalization, and MMR helpers**: Simplifies vector operations. - **Representation-independent vector conversion and mean pooling**: Handles various input formats. - **Optional native GPU execution**: Accelerates search using wgpu. - **Top-level `Vettore.*` API**: Simplifies integration with Elixir systems. - **Compatibility with `Vettore.new/0`**: Supports both new and old APIs. ### Tech Stack: - **Elixir**: The primary language for the project. - **Rust**: Used for acceleration in specialized operations. ### Why It Matters: - **Efficiency**: Keeps data in ETS for fast

Python 25 stars

ericmillsio/whiplash

Serverless, lightweight, and fast vector database on top of DynamoDB

Overview: Whiplash is a lightweight vector store built on top of AWS DynamoDB, designed for easy use, scalability, and fast performance. Key features: - Whiplash Serverless API - Whiplash Client SDK - Whiplash CLI Tech stack: Python Why it matters: Whiplash simplifies the development of vector databases by providing a lightweight, scalable, and easy-to-use solution that can be easily integrated into existing AWS infrastructure. Repository: ericmillsio/whiplash Primary language: Python GitHub stars: 25 GitHub description: Serverless, lightweight, and fast vector database on top of DynamoDB

TypeScript 25 stars

gezilinll/Proteus

The AI-native editor for multimodal creation

Overview: Proteus is an AI-native editor core designed for multimodal creation, focusing on an open-source project that aims to evolve into a platform supporting user-defined LLM, MCP, Skill, and Plugin capabilities. Key features: - Supports AI-driven agents, LLMs, MCPs, and Skills. - Enables seamless integration of various AI tools and capabilities. - Provides a secure environment for AI execution, rather than relying on a large number of AI features. - Focuses on core editor functionality, including basic drawing, text, and image elements, as well as interactive tools and layer management. Tech stack: TypeScript Why it matters: Proteus is a crucial component for developers and designers who need to create and edit multimodal content efficiently, ensuring a secure and efficient workflow while integrating AI-driven tools seamlessly. Repository: gezilinll/Proteus Primary language: TypeScript GitHub stars: 25 GitHub description: The AI-native editor for multimodal creation README excerpt: <div align="center"> # Proteus Editor **面向 AI 集成的编辑器核心，公开构建中** *An AI-ready editor core, built in public* [English](./README.en.md) · [在线体验](https://proteus.gezilinll.com/) · [路线图](./plans/phase-step-overview.md) · [Spec 工作流](./specs/README.md) · [技术文章](./articles/) [![License: MIT](https://img.shields.io/badge/License-MIT-blue.svg

TypeScript 25 stars

gvkhosla/pi-tinker

Fine-tune open-source models with Tinker from inside Pi — managed improve loops, data prep, evals, smoke tests, deploy snippets, and checkpoint chat.

Overview: pi-tinker is a tool designed to fine-tune models on Tinker from inside Pi, managed by Tinker itself. It includes data preparation, Python script writing, small training jobs, and checkpoint evaluation, all under the Tinker framework. Key features: - Data preparation and editing - Python script creation - Small training job execution - Smoke tests and checkpoint evaluation - Model fine-tuning with Tinker - Support for various model types and contexts Tech stack: TypeScript Why it matters: pi-tinker simplifies the process of fine-tuning models on Tinker, making it easier for developers to experiment and improve their models without needing to set up their own infrastructure. Repository: GitHub - gvkhosla/pi-tinker Primary language: TypeScript GitHub stars: 25

Python 25 stars

hellochick/text-to-image

Re-implementation of https://github.com/zsdonghao/text-to-image

Overview: This project re-implementation of the text-to-image algorithm, focusing on the Kaggle competition "Reverse Image Caption". It utilizes the Generative Adversarial Text-to-Image Synthesis method and StackGAN for advanced image generation. Key features: - Implementing the text-to-image algorithm for the Kaggle competition - Utilizing the Generative Adversarial Text-to-Image Synthesis method - Incorporating the StackGAN architecture for enhanced image generation Tech stack: Python Why it matters: This project showcases advancements in text-to-image technology, demonstrating the capability to generate realistic images from textual descriptions, which has applications in various fields including art, design, and creative industries. Repository: hellochick/text-to-image Primary language: Python GitHub stars: 25 GitHub description: Re-implementation of https://github.com/zsdonghao/text-to-image

Python 25 stars

hiper2d/ai-llm-playground

Chatbot web-applications with LLM, OpenAI API Assistants, LangChain, vector databases, and other AI stuff

### Overview: This project explores the integration of OpenAI Assistant API with LangChain and vector databases, focusing on restaurant recommendations, web scraping, and conversational AI. ### Key Features: - Restaurant Advisor (OpenAI Assistant version) - Web Scraper based on Vision API - Chat with PDF documents using OpenAI Assistant API - AI Girlfriend - Chat with Multiple Documents (Outdated Langchain version) - Setup ### Tech Stack: Python ### Why It Matters: This project showcases the capabilities of integrating cutting-edge AI tools like OpenAI Assistant API, LangChain, and vector databases, providing a practical application of these technologies in a conversational AI context.

Python 25 stars

isenlink/thor-fp8-llm

### Overview: 本项目旨在为 NVIDIA DRIVE Thor 域控板上的本地大模型部署提供详尽的文档和实测数据，包括从零部署到生产定稿的全过程。通过交叉编译、模型转换、推理调优等环节，确保在 DRIVE Thor 上成功部署和优化 LLM（如 llama.cpp）。 ### Key Features: - **部署文档**：详细说明了 DRIVE Thor 上的 LLM、H3、Image 和 vLLM 四线部署的共同基础。 - **底座二进制**：提供 CUDA 13.2 共用底座的模型，适用于后续所有部署。 - **生产部署线**：2026-10-25，展示了 vLLM 0.32 / SM101 独立测试和 MTP 与并发验证的生产部署线。 - **历史数据解读**：与下方 llama.cpp 历史数据分别解读，提供详尽的实测结果和优化路径。 ### Tech Stack: 使用 Python 编程语言，结合 CUDA 13.2 作为底座技术栈。 ### Why It Matters: 本项目为 DRIVE Thor 上的本地大模型部署提供了详尽的实测数据和优化路径，填补了 DRIVE Thor 民间部署资料的空白，有助于节省其他开发者的时间和精力。

Python 25 stars

iut62elec/Soccer-Highlight-Generator-with-GenAI

Automate the creation of soccer match highlights with the power of Generative AI and AWS. This solution leverages AWS Bedrock (Anthropic’s Claude 3 Sonnet model), AWS MediaConvert, Lambda, Step Functions and other AWS services to identify and compile exciting game moments without manual editing.

### Overview: This project automates the creation of soccer match highlights using Generative AI and AWS services. It leverages AWS Bedrock (Anthropic's Claude 3 Sonnet model), AWS MediaConvert, Lambda, and Step Functions to identify and compile exciting game moments without manual editing. ### Key Features: - Automatically generates soccer highlight videos. - Uses AWS Bedrock for AI model access. - Integrates AWS MediaConvert for video processing. - Utilizes Lambda and Step Functions for orchestration. - AWS services include S3, DynamoDB, and SQS. ### Tech Stack: Python ### Why It Matters: This solution accelerates the process of creating high-quality soccer highlight videos, reducing manual editing and ensuring consistent results. It streamlines the workflow by automating the identification and selection of game moments, making it easier for teams and fans to access exciting moments from matches.

Python 25 stars

jiahongsigma/Efficient-LLM-Inference-Serving-Systems

Why is LLM inference slow — and how do you make it fast? A hands-on, first-principles course: roofline → KV cache → quantization → parallelism → vLLM/SGLang, with GPU labs on open models.

Overview: This repository is a comprehensive, hands-on course designed to explore and optimize the performance of Large Language Models (LLMs) in production settings. The course delves into various techniques such as batching, KV caching, quantization, parallelism, speculative decoding, and structured serving, aiming to maximize the efficiency of GPU usage. Key features: - Batch processing - KV cache management - Quantization - Tensor parallelism - Long-context serving - Speculative decoding - Structured and constrained decoding - Tool-use and agentic serving - Frameworks, API, and observability - Evaluation methodology Tech stack: Python Why it matters: Understanding and optimizing LLM inference and serving systems is crucial for developers aiming to improve the performance and efficiency of their models, especially in terms of GPU utilization. By mastering the techniques discussed in this course, developers can significantly enhance the speed and resource efficiency of their LLMs, leading to better overall system performance and faster response times. Repository: jiahongsigma/Efficient-LLM-Inference-Serving-Systems Primary language: Python GitHub stars: 25 GitHub description: A hands-on, first-principles course on how LLMs are actually served in production — from the roofline up to vLLM & SGLang, with GPU labs on open models and every result reproducible.