📦 tonyd2wild/glm-5.2-quanttrio-200k-4x-dgx-spark

glm-5.2-quanttrio-200k-4x-dgx-spark

You can now serve the massive, unpruned GLM-5.2 model with all 256 experts intact at an incredible 200,000-token context window. This repository provides a complete, fully replicable recipe to run this high-fidelity setup across a four-node cluster of NVIDIA DGX Spark systems. By opting for a clever mixed precision checkpoint, the recipe fits the entire model into memory with actual room to breathe, completely avoiding the quality compromises of pruned models. It utilizes speculative decoding and optimized Triton kernels to achieve a blazing-fast speed of nearly thirty tokens per second. It is a massive win for anyone wanting top-tier open-source intelligence running locally at scale.

🆔 @hackernewsgithubprojects

📦 exxeta/exxperts

Exxperts AI

Control exactly what your AI remembers by running a local-first agentic runtime right on your computer. While typical AI tools silently build profiles or forget your conversations between sessions, this system keeps your persistent workspace rooms, knowledge bases, and files entirely on your own disk as plain files. The most exciting part is the approval-gated memory checkpoint system, meaning the AI must ask your permission before saving any new facts, decisions, or preferences about you. You can easily interact with your persistent rooms through a web dashboard or jump straight into your terminal with a command-line interface. It is the perfect way to get smart, personalized assistance without giving up...

📰 https://news.ycombinator.com/item?id=48884002

🆔 @hackernewsgithubprojects

📦 rasbt/llm-architecture-gallery

LLM Architecture Gallery

Most major language models are actually built on slightly different blueprints, and you can compare them all in one visual dashboard. The llm-architecture-gallery is a public metadata hub that powers a live visual gallery tracking the exact architectural differences between famous AI models. Instead of digging through endless research papers to find hidden technical details, this project catalogs the precise knobs, layers, and settings that make each model unique. It is a fantastic way to see how AI design is evolving without drowning in academic jargon. Check out the gallery to see how your favorite model is actually put together.

🆔 @hackernewsgithubprojects

📦 jinwoolee1230/poli

poli

Turn sparse, messy lidar point clouds into rich geometric understanding without relying on labeled data. Instead of using handcrafted rules or heavy supervised learning to clean up noisy 3D sensor scans, this self-supervised tool predicts a mathematical ellipsoid for every single point to map the surrounding surface shape. This means robots can understand their environments, track their own movements, and estimate object positions with much higher accuracy. It plugs directly into existing tracking and mapping systems without needing any changes to your neural network architecture. By modeling raw data as smooth mathematical shapes, you get clean, reliable 3D geometry instantly.

🆔 @hackernewsgithubprojects

📦 igorbarshteyn/jlens-gguf

jlens-gguf

You can now visually track and actively edit a language model's internal thoughts in real time using a standard quantized GGUF file. The jlens-gguf repository is an interactive visualizer and steering tool built directly on llama.cpp that lets you peer into the residual stream of a model at every layer. By mapping out how concepts form, it lets you pinpoint where the AI gets confused and actually edit those activations on the fly, allowing you to swap concepts or suppress specific tokens to instantly steer the generated text. It is an incredibly clever way to dissect and control local models with zero hassle.

🆔 @hackernewsgithubprojects

📦 sizhe-li/vera

Vera

Vera is the video-to-action model that finally lets robots control themselves by dreaming up their own future. Instead of trying to teach a robot physical actions directly, this clever setup lets a standard video-generating AI imagine a video of the task being completed successfully. Then, a specialized translating model reads that imagined video and calculates the exact physical joint movements needed to make that dream a reality. It completely separates dreaming from doing, making it incredibly easy to train robots for complex tasks without needing a massive supercomputer. You can even watch the robot think in real time through a live visual dashboard.

🆔 @hackernewsgithubprojects

📦 luckeyfaraday/athena-loops

Athena Loops

While AI coding agents are incredibly powerful, letting them run wild without supervision often leads to broken code and wasted API tokens. Athena Loops changes the game by turning the orchestrator, worker, and reviewer process into a deterministic control loop written in pure Python code. Instead of relying on a prompt to hopefully self-correct, this smart harness splits your goal into subtasks, delegates them to headless coding agents, and passes the output through a strict review gate that loops until success criteria are met. Because the control flow is handled by code rather than unreliable prompts, it prevents infinite agent loops using built-in budget guards. It even runs as an...

🆔 @hackernewsgithubprojects

📦 endee-io/endee

Endee Vector Database

Endee is a dedicated vector database written in C++ that can index and search up to one billion vectors on a single machine. Built to handle heavy AI retrieval workloads and RAG pipelines, it solves the problem of slow search times by optimizing its indexing directly for modern hardware. It compiles down to target specific CPU instructions like AVX512 or ARM's NEON to extract absolute peak performance. The coolest part is that it supports hybrid search, combining dense vector similarity with sparse exact matches and real-time metadata filtering. It is a incredibly fast, self-contained engine for giving your AI applications lightning-fast long-term memory.

🆔 @hackernewsgithubprojects

📦 rit3zh/expo-morphing-menu

expo-morphing-menu

Build sleek, fluid user interfaces in React Native by turning your standard chat composer into an interactive menu. Instead of letting a traditional pop-up sheet block your app design, expo-morphing-menu uses a spring physics engine to morph the input bar itself. The composer dynamically expands in height, width, and position to become the menu background. It stays perfectly aligned with the mobile keyboard while navigating multiple panels like photos, files, or custom integrations. It offers mobile developers a clean way to handle text inputs, media attachments, and options panels inside a single, unified interface.

🆔 @hackernewsgithubprojects

📦 hugecatlab/chattutor

ChatTutor

ChatTutor is the interactive learning assistant that finally gives your AI tutor a digital blackboard to sketch on. While standard chatbots get stuck explaining complex math and science using only dry text, ChatTutor gives the AI hands-on control over a dynamic math canvas and interactive mindmaps. When you ask a question, the AI actually draws out formulas, plots geometric shapes, and structures ideas visually, just like a real teacher writing on a whiteboard. It completely changes how you study technical topics. Check out the project to host your own visual AI classroom and make learning feel alive.

🆔 @hackernewsgithubprojects

📦 xiangyu-cas/codex-canvas

Codex Canvas

Edit images on a local infinite canvas directly inside your Codex conversation. This tool automatically pulls your generated visuals into a dedicated sidebar panel, keeping your workspace organized by project thread. The coolest feature is how you can isolate layers. It separates flat images into independent, movable elements like background, text, and objects, completing the hidden background behind them. You can rearrange these layers on the fly and export the entire canvas group as a Photoshop-ready PSD file. It makes visual asset organization incredibly easy. Give it a spin to bring layered, canvas-style image editing straight into your chat workflow.

🆔 @hackernewsgithubprojects

📦 adamlyttleapps/claude-skill-app-onboarding-questionnaire

claude-skill-app-onboarding-questionnaire

Build a high-converting, psychology-backed onboarding flow for your mobile app in minutes. The claude-skill-app-onboarding-questionnaire is a clever extension for your AI assistant that scans your codebase, figures out your app's core value, and automatically codes a gorgeous, multi-screen onboarding sequence. The absolute coolest part is that it doesn't just write basic questions. It actually builds a fully functional mini-demo of your core app experience right inside the onboarding, giving users a taste of your product before they even hit the paywall. It handles all the layout and copywriting automatically so you can watch your conversions skyrocket.

🆔 @hackernewsgithubprojects

📦 jino-rohit/ml-systems-notes

ml-systems-notes

You can calculate the exact computing cost of a LLaMA 3 forward pass down to the individual floating-point operation just using pen and paper. That is just one of the cool secrets tucked away in ml-systems-notes, a personal notebook packed with the actual mechanics of machine learning engineering. Instead of just using frameworks, this repository breaks down how complex tasks actually run on hardware, from the exact math of model quantization to the way CUDA graphs eliminate CPU bottlenecks during model decoding. It is basically the ultimate cheat sheet for anyone who wants to understand how massive models actually run under the hood. Check it out and write your own...

🆔 @hackernewsgithubprojects

📦 gajesh2007/ai-trading-agent

AI Trading Agent

This open-source AI trading agent completely automates cryptocurrency trading on the Hyperliquid decentralized exchange by running real-time market data directly through advanced language models like GPT-5 and DeepSeek R1. The system runs in a continuous loop, analyzing technical indicators and live price action to make autonomous trading decisions and place orders. Instead of relying on rigid, pre-programmed rules, the agent uses its reasoning capabilities to evaluate market sentiment, execute buy or sell signals, and manage active positions with custom take-profit and stop-loss targets. It is a fascinating look at how modern language models can handle complex, real-world execution environments in real-time.

🆔 @hackernewsgithubprojects

📦 jangviktor-web/nihaixia

Nihaixia

Bring the legendary clinical wisdom of a master practitioner straight into your AI workspace. This project distills over two thousand pages of classic Chinese medicine teachings, diagnostic formulas, and real-world clinical cases into an interactive agent skill. You can prompt the assistant to analyze health patterns and guide remedy selection using historical methodologies, all rendered in the practitioner's signature, plain-spoken teaching style. It is an incredibly detailed blueprint for anyone curious about how traditional knowledge can be structured for modern software. Explore the repository today to see this unique logic system in action.

🆔 @hackernewsgithubprojects

📦 hugo-dz/spritefusion-pixel-snapper

Sprite Fusion Pixel Snapper

Generative AI is notoriously terrible at understanding grid-based pixel art, usually producing messy images with drifting resolutions and inconsistent color palettes. That is where spritefusion-pixel-snapper comes in, using Rust and WebAssembly to automatically snap those chaotic, off-grid pixels back into a perfect, clean layout. It detects the native pixel dimensions, aligns the image to a consistent grid, and quantizes the colors into a strict, beautiful palette. Whether you are fixing AI-generated game assets or cleaning up procedural 2D tilemaps, it gives you crisp results instantly. Run it directly in your browser or batch-process folders of images offline using the command line.

🆔 @hackernewsgithubprojects

📦 cosmtrek/mindwalk

Mindwalk

Mindwalk is the visualization tool that replays AI coding sessions on a three-dimensional map of your codebase. When an AI agent works on your files, reading raw log files makes it incredibly difficult to tell what the assistant actually did or how it explored. This tool builds a deterministic, dark-mode city map of your repository and traces the agent's actions as moving light, turning searches, reads, and edits into glowing paths you can inspect at a glance. It runs completely locally on your machine to analyze your sessions. Check it out to see exactly how your AI assistant thinks.

📰 https://news.ycombinator.com/item?id=48878682

🆔 @hackernewsgithubprojects

📦 charlie947/social-media-skills

Social Media Skills

Put your social media growth on autopilot by installing the exact AI agent workflows used to reach over three hundred and fifty thousand followers. This repository lets you clone seventeen specialized markdown skills directly into your Claude desktop app or workspace. It completely replaces generic AI writing with hyper-focused tasks, from reverse-engineering viral short-form video scripts to generating data-driven LinkedIn post scorecards and custom graphics. It solves the headache of bland, robotic AI content by grounding every output in your unique, pre-analyzed voice profile. Download social-media-skills to turn your favorite LLM into an elite, multi-platform media team.

🆔 @hackernewsgithubprojects

📦 jakubantalik/transitions.dev

transitions.dev

Build sleek, animated web interfaces in seconds by dropping ready-to-use, professional motion patterns straight into your codebase. With transitions.dev, you get an interactive playground of essential CSS transitions, like smooth card resizing, springy notification badges, and origin-aware dropdowns. Each interaction comes with a self-contained, copy-pasteable CSS snippet wrapped in reduced-motion guards. Even cooler, it acts as an installable agent skill for your AI coding tools, letting Claude or Cursor apply these polished motion tokens directly inside your files. It is the ultimate shortcut to making your website feel tactile and alive without fighting complex animation libraries.

🆔 @hackernewsgithubprojects

📦 shadcnstore/shadcn-dashboard-landing-template

shadcn-dashboard-landing-template

Build a fully featured web application and its marketing homepage at the exact same time without starting from scratch. The shadcn-dashboard-landing-template repository provides a complete, production-ready admin panel alongside a beautiful marketing landing page, working out of the box with both Vite and NextJS. Instead of spending days wireframing and styling, you get thirty pre-built pages, interactive charts, and complete app layouts for mail, chat, and tasks. The standout feature is a live theme customizer that lets you tweak colors and sidebar layouts in real-time. Grab this template to jumpstart your next SaaS or AI startup instantly.

🆔 @hackernewsgithubprojects