Top 10 Must‑See Open‑Source Projects on GitHub This Week

01. Omarchy – A Ready‑to‑Use Linux Distro

Omarchy, launched by DHH, is an open‑source Linux distribution based on Arch Linux that uses the Hyprland desktop for automatic window tiling. It ships with pre‑configured themes, shortcuts, common software, and system settings, providing a unified desktop environment suitable for both office work and development. The ISO includes Neovim, Chromium, Obsidian, LibreOffice, Kdenlive, OBS Studio, Docker, and various programming‑environment installers. Installation enables full‑disk encryption and wipes the target drive, saving developers the effort of configuring Arch from scratch.

02. Semantica – AI Agent Knowledge‑Graph Infrastructure

Semantica is an open‑source graph platform for AI agents that transforms scattered documents, web pages, and databases into a structured knowledge graph. It automatically extracts entities, companies, events, and their relationships, allowing agents to understand not only similar content but also the provenance, connections, and potential conflicts of information. The system records audit trails of AI decisions, supports rule‑engine checks, and can export data in JSON, CSV, or RDF formats, making it suitable for compliance‑heavy domains such as finance, healthcare, and law.

开源地址：https://github.com/semantica-agi/semantica

03. OpenLogi – Open‑Source Logitech Peripheral Manager

OpenLogi offers a login‑free, open‑source alternative to Logitech Options+. It manages compatible Logitech mice, keyboards, and UVC cameras, displaying battery levels and allowing users to remap mouse buttons, keyboard F‑keys, gestures, DPI, and SmartShift scrolling. The tool also supports RGB lighting for certain keyboards and focus, exposure, brightness, and white‑balance adjustments for cameras. It runs on macOS, Linux, and Windows without collecting telemetry data.

开源地址：https://github.com/AprilNEA/OpenLogi

04. DeepSeek Harness & Cordis – Plugin Framework for AI Agents

DeepSeek Harness uses Cordis as its underlying plugin engine, providing a modular architecture where each module declares required capabilities, startup timing, and inter‑module communication. This design enables developers to add new tools, replace models, or adjust abilities without modifying the entire system. Cordis also handles cleanup of listeners and resources when modules are removed or reloaded, ensuring flexibility and extensibility.

开源地址：https://github.com/cordiverse/cordis

05. OpenViking – Context Database for AI Agents

OpenViking, an open‑source project from Volcano Engine, serves as a unified context database for AI agents, storing memories, project materials, and skills. It organizes data in a hierarchical file system, allowing agents to query using commands like ls, tree, and find. The system stores information in three layers—summary, overview, and full content—to reduce irrelevant token consumption. After each session, it extracts user preferences and task experience into long‑term memory, supporting tools such as Claude Code, Codex, Cursor, and OpenClaw. The database is persisted as Markdown files in a Git repository, enabling search, backup, and offline operation.

开源地址：https://github.com/volcengine/OpenViking

06. AI Memory – Long‑Term Memory System for AI Coding Assistants

AI Memory records filtered prompts, tool operations, and session boundaries via hooks, consolidating valuable content into a project‑specific Markdown wiki. It captures architecture decisions, technical trade‑offs, failures, and pending issues, enabling a new agent to quickly understand progress without re‑explaining the entire background. The system supports handoff between agents—for example, after exiting Claude Code, launching Codex in the same directory allows the new agent to receive the previous session’s summary, to‑do list, and handover information. It integrates with Claude Code, Codex, Cursor, Gemini CLI, OpenCode, and stores data in a Git repository for easy backup and offline use.

开源地址：https://github.com/akitaonrails/ai-memory

07. llmfit – Automatic Local Model Selection Tool

llmfit is a terminal tool that automatically detects a computer’s RAM, CPU, GPU, VRAM, and backend, then evaluates hundreds of models with various quantization schemes. It predicts which models fit into memory, estimates runtime speed, and assesses model quality and context length, preventing users from downloading large models that their hardware cannot handle. The tool offers an interactive CLI, supports multi‑GPU and mixed‑expert models, and can work with Ollama, llama.cpp, MLX, Docker Model Runner, and LM Studio.

开源地址：https://github.com/AlexsJones/llmfit

08. oMLX – Apple Silicon Local Inference Service

oMLX provides a complete local inference stack for Apple Silicon, accessible from the macOS menu bar. It supports continuous batching and concurrent requests on a single machine. A hierarchical KV cache keeps hot data in RAM and offloads cold data to SSD, allowing cache restoration after a reboot. oMLX can serve text LLMs, vision‑language models, OCR, embeddings, and rerankers. Its management panel displays runtime metrics, enables model downloads, benchmark runs, and lets users set aliases, context windows, TTL, and sampling parameters; high‑frequency models stay resident, while low‑frequency models are evicted using an LRU policy.

开源地址：https://github.com/jundot/omlx

09. Unsloth – Fast, Low‑VRAM AI Model Training Toolkit

Unsloth is an open‑source AI model training suite with 74 k stars that tackles slow fine‑tuning, high VRAM usage, and complex configuration. Official benchmarks show up to a 2× speed‑up and roughly 70 % VRAM reduction, enabling developers with consumer‑grade GPUs to fine‑tune models. It offers desktop, web‑UI, and code interfaces, covering data processing, LoRA/QLoRA fine‑tuning, model export, and local deployment. While performance varies with model size, hardware, and training strategy, clean and stable data remains essential for good results.

开源地址：https://github.com/unslothai/unsloth

10. Needle 2 – 14 MB Tool‑Calling Model for Edge Devices

Needle 2 is a 45 M‑parameter open‑source model designed for tool calling, device control, and structured information extraction. Packaged with its runtime engine, the model is only about 14 MB, and a full session consumes roughly 28 MB of RAM, making it suitable for phones, wearables, smart home devices, and small robots. It selects appropriate functions from natural‑language prompts, fills parameters, and returns JSON results. After the initial download, inference runs completely offline.

开源地址：https://github.com/cactus-compute/needle