Runnable, documented projects worth studying and extending.
Learning and educationBuild Your Own XPWRealize the technology you love from scratch! A curated collection of tutorials covering dozens of areas including 3D renderers, databases, compilers, operating systems, and other areas, showcasing the best practices of the Feynman learning method.
Efficiency toolsLightpanda BrowserPWA headless browser designed specifically for AI and automation, written in Zig, extremely lightweight and high-performance, suitable for web scraping and agent web interaction scenarios.
AI and AgentsDeerFlow - ByteDance's super agent frameworkPWByteDance's open-source super agent framework integrates sandboxes, memory, tools, and sub-agents, capable of handling complex tasks such as research, coding, and content creation
Efficiency toolsMarkItDown - Microsoft file conversion toolPWMicrosoft's open-source file-to-Markdown tool, supporting 20+ formats including PDF, Word, Excel, PPT, images, and audio, specifically designed for LLM text analysis pipelines. Provides MCP Server integration, enabling seamless integration with LLM applications such as Claude Desktop.
OthersChainlitPWQuickly build a Python framework for AI chat applications, supporting streaming, multimodality, and human loops
SafetyTrivyPWOne-stop security scanner: Detection of vulnerabilities, configuration errors, and key leaks in containers, K8s, code repositories, and cloud environments—an essential tool for DevSecOps.
Efficiency toolsStable Diffusion WebUIPWThe browser interface of Stable Diffusion supports all functions including txt2img, img2img, ControlNet, and more
OthersBitNetPWMicrosoft's official 1-bit LLM inference framework significantly reduces deployment costs and hardware requirements for large models, supporting efficient execution of large models on CPUs
AI and AgentsMiroFishPWA simple and universal swarm intelligence engine that can predict everything. Based on swarm intelligence algorithms, it supports various prediction scenarios such as stock market, weather, and sports, implemented in Python, and today GitHub has 399 stars.
Efficiency toolsSillyTavernPWAI role-playing chat frontend for advanced users, supporting multiple LLM backends and rich customization options
OthersApollo-11 source codePWThe original source code for NASA's Apollo 11 lunar module navigation computer (AGC), including the command module and lunar module modules. A historic open-source project with 65k+ stars.
Efficiency toolsMinerUPWA one-stop open-source data extraction tool that converts PDFs to Markdown and JSON, supports OCR
AI and AgentsClaude officially uses a recipe collectionPWAnthropic's official Claude recipe collection covers scenarios such as agent construction, tool calls, and multimodal applications.
Efficiency toolspromptfoo - Prompt review and red team testingPWCLI tools for evaluating and red-teaming LLM applications. Supports automated prompt evaluation, model comparison, and security vulnerability scanning, with integration of CI/CD. MIT is open source, supporting production applications for tens of millions of users.
OthersOpenAI SkillsPWOpenAI's official Codex Agent Skills catalog defines a set of reusable instructions, scripts, and resource pack standards, enabling AI Agents to automatically discover and execute specific tasks. Supports custom skill creation and distribution.
Efficiency toolsMiroFish - Swarm Intelligence Prediction EnginePWAn AI prediction engine based on multi-agent technology builds a high-fidelity parallel digital world, enabling thousands of agents to freely interact and evolve in society, enabling precise prediction of future trends. Supports various prediction scenarios such as financial signals, public opinion events, and novel endings.
Efficiency toolsAn AI tool that can automatically generate hand-drawn storyboards for a given video link: clipsketch-ai, and automatically adapt it to self-media platform copywriting for one-click promotionPWclipsketch-ai is an AI tool that can automatically generate hand-drawn storyboards based on video links and adapt copywriting for self-media platforms. It supports batch generation and optimized storyboarding, allowing users to upload custom characters to fit into scenes, and supports parsing video links from Bilibili and Xiaohongshu
Efficiency toolsMarkerPWFast and accurate PDF to Markdown tool, supporting tables, formulas, and code blocks
Efficiency toolsZhipu just released another OCR: GLM-OCR, a PDF at 1.86 pages per second, can run 2,000 A4 scans for 1 yuan, 0.9B, and supports edge deploymentPWGLM-OCR is an efficient OCR tool that processes PDF documents at speeds of 1.86 pages per second, capable of processing about 2,000 A4 scans for 1 yuan, at only one-tenth the cost of traditional OCR. It supports handwriting, complex tables, and multilingual mixed layouts, suitable for high concurrency and edge deployments, and scored an impressive 94.6 on OmniDocBench V1.5. It can be deployed via vLLM, SGLang, and Ollama, reducing inference latency and computing power overhead
OthersRuViewPWWiFi DensePose: Real-time human posture estimation, vital sign monitoring, and presence detection using ordinary WiFi signals, completely eliminating the need for a camera. Built on Rust, high performance and low latency.
OthersAlibaba has just open-sourced its latest 80B activated 3B MoE programming model: Qwen3-Coder-NextPWAlibaba has open-sourced its latest 80B activated 3B MoE programming model, Qwen3-Coder-Next, excelling at long-duration and multi-tool programming tasks, with a resolution rate exceeding 70%. This model can understand requirements, write code, interact with the environment, and complete tasks, supporting local IDE plugins, CLI agents, and enterprise intranet deployment
OthersAnt Health and others have just open-sourced a medical language model: AntAngelMedPWAntAngelMed, an open-source medical language model from Ant Health, is based on the Ling-flash-2.0 MoE architecture, featuring 100B total parameters and 6.1B activation parameters, inference speeds exceeding 200 tokens/s, and support for 128K contexts. Ranked first on the HealthBench benchmark, after three stages of training, possesses strong medical knowledge and complex diagnostic and treatment reasoning skills, suitable for analyzing complex cases and daily health Q&A
OthersNVIDIA has released: Cosmos Reason 2, the latest inference visual language model specializing in physics AI, with a context length of 256KPWNVIDIA has launched Cosmos Reason 2, a cutting-edge inference visual language model focused on physical AI, supporting context lengths up to 256K. This model enhances spatiotemporal understanding and timestamp accuracy, enabling 2D/3D point positioning, boundary box coordinates, trajectory data, and OCR, outputting robot movements and movement trajectories, suitable for video analysis, data annotation, and security inspection applications. Available in two models: 2B and 8B
OthersAlibaba has just released its latest models: Qwen3-VL-Embedding and Qwen3-VL-Reranker, with multimodal information search at its corePWAlibaba has launched the Qwen3-VL-Embedding and Qwen3-VL-Reranker models, supporting multimodal information search, including text-based image search and image search. This combination improves search efficiency through a two-stage retrieval process of Embedding and Reranking, supports over 30 languages, and Embedding-8B performs excellently on MMEB-V2