📦 divagr18/qwenpu

The LLM That Runs Inside An LLM

QwenPU is the language model that runs a tiny computer inside its own brain. It takes a popular four billion parameter model and compiles a full RISC-V processor and vector unit directly into its neural network weights. The result is a machine that can execute real C programs without any external hardware. It actually runs Andrej Karpathy’s tiny story writer, which uses this internal computer to generate text. Every step is calculated by the model’s own layers, making it a computer inside a computer. It is still very large and slow, but the concept is wild. A single neural network that thinks and computes at the same time.

🆔 @hackernewsgithubprojects

📦 rowow-llc/open-sourced-off-the-shelf-ion-exchange-membrane

Build a $1 Ion Exchange Membrane at Home

Fabricate a functional ion exchange membrane for under one dollar per square foot using materials found in any hardware store. This project provides the complete open-source recipe and patent to create this critical component for electrochemistry without specialized suppliers or industrial equipment. By mixing pulverized water softener resin with standard PVC plastic and common solvents, you can cast your own membrane that filters ions as effectively as expensive commercial alternatives. This dramatically lowers the barrier to entry for recovering precious metals from e-waste or building flow batteries. It turns a costly industrial monopoly into a DIY hobbyist project.

🆔 @hackernewsgithubprojects

📦 jaydenteoh/soft-mtp

Soft Multi Token Prediction Speeds Up AI

Soft Multi Token Prediction is the training method that finally makes language models generate text faster without sacrificing quality. The repository provides the official code for a technique that fixes a hidden flaw in how neural networks are taught to predict multiple words at once. Standard methods force the model to treat its internal calculations as separate, which wastes a massive amount of computational power. By using a clever mathematical trick to share information between the model's internal layers, this project allows the network to draft several words simultaneously.

🆔 @hackernewsgithubprojects

📦 mikeparcewski/wicked-studio

Wicked Studio: The Control Room for Coding Agents

Wicked Studio is the control room that lets you watch your coding agents work in real time. It solves the problem of flying blind when AI writes your code, because here you start a job with a simple chat, then sit back and observe the agents working inside safe, isolated environments. The coolest part is that you don't have to trust the process blindly; the interface shows you exactly what the agent did, so you can review the evidence and approve only what actually matters before it touches your real files. This turns a mysterious black box into a transparent workspace where you stay in charge of the final result.

🆔 @hackernewsgithubprojects

📦 ethbak/icon-composer-mcp

Icon Composer: Liquid Glass Icons via AI

Icon Composer is the tool that turns a flat logo into a living Apple Liquid Glass icon without opening a single design app. It speaks directly to AI assistants, letting you build, tweak, and render high-fidelity icon bundles right from your chat. This solves the tricky problem of matching Apple’s strict design standards by automating the complex rendering pipeline. It is the perfect bridge for developers who want pixel-perfect, store-ready assets without the usual design friction. If you build for Apple platforms, this is your shortcut to professional-looking visuals.

🆔 @hackernewsgithubprojects

📦 synchopate/nemotron-30b-cpu-256k

Nemotron 30B CPU 256K

A thirty billion parameter language model now runs entirely on a standard laptop processor without any graphics card, achieving speeds that feel instant for interactive coding. This repository explains how to run Nemotron thirty B on a central processing unit using a hybrid architecture that replaces most attention layers with state space models. This design drastically reduces memory usage, allowing a massive two hundred and fifty six thousand token context window to fit comfortably in system memory. By combining this efficient model with low level system optimizations, the setup handles large context windows at speeds rivaling dedicated hardware.

🆔 @hackernewsgithubprojects

📦 bendbss/riposte-frontline-source

Riposte: Frontline

Riposte: Frontline is the open-source first-person shooter that builds every single 3D object from pure math code instead of importing standard 3D model files. You might not expect a full-featured game with dynamic water and complex vegetation to run so smoothly in a plain web browser, but this project proves the engine can handle it all using just JavaScript. It is completely free to download, run locally, or remix for your own projects under a permissive license. The codebase is massive and runs entirely in your browser without any install. If you want to see how game developers build worlds from scratch, grab the source and start exploring.

🆔 @hackernewsgithubprojects

📦 agent-passport-system/agent-passport-system

Agent Passport: Giving AI Agents a Verifiable Identity

Give any AI agent a cryptographically signed identity that proves exactly what it is allowed to do. The agent passport system is an open protocol and SDK that stops agents from acting on vague trust by checking their permissions before they perform any task. Think of it as a digital badge that a system verifies in milliseconds, ensuring an agent only does what it was specifically hired to do. If the agent tries to step out of bounds, the system blocks the action instantly and leaves a permanent, signed record of the attempt. This gives developers a way to build autonomous tools without handing over the keys to the unknown.

🆔 @hackernewsgithubprojects

📦 deibler/openskycam

Point a camera at the sky and let AI find the meteors

Point a cheap camera at the sky and let software hunt for meteors. Open Sky Cam turns a standard internet camera into a twenty-four hour recording system that automatically spots planes, falling stars, and moving objects without sending your private footage to a remote server. Instead of just storing video, the system actually watches the sky for you using AI detection models to find and label things flying past. It separates the interesting stuff, like a falling meteor, from the background noise, so you can browse specific events instead of scrubbing through hours of empty sky.

🆔 @hackernewsgithubprojects

📦 bubblebrainbb/bb-video

BB Video Makes Movies From Code

BB Video is the agent skill that turns plain text into fully animated videos without a single frame of stock footage. You just tell the agent what story you want and pick one of twenty-five visual styles, like shadow puppets or pixel art. The magic happens in the background: it writes code that draws every character, sound effect, and musical note from scratch, then renders the whole thing frame by frame in a browser. It even checks its own work to make sure nothing overlaps or looks broken. You do not need to know how to animate or edit video.

🆔 @hackernewsgithubprojects

📦 ysuito/rabbitwall

RabbitWall: The Kill Switch for Your AI Network

RabbitWall is the network gateway that lets you watch and control your AI traffic. It sits between your local network and your AI servers, giving you a visual map of every connection. You can see exactly which domains your AI is contacting and where they are located. The standout feature is a simple switch that instantly cuts off all internet access. This stops data from leaving your machines, even if they are already connected. It is perfect for anyone running sensitive AI tools who needs strict control. You get total visibility and instant power to disconnect. Keep your data secure and know exactly what is happening.

🆔 @hackernewsgithubprojects

📦 alibaba-damo-academy/rynnunify

RynnUnify: The AI That Dreams Before It Acts

RynnUnify is the embodied AI system that actually imagines the future before it moves its hands. Instead of just recognizing objects, it generates a short video prediction of how a scene will change, then uses that visual preview to plan its next step. This allows the robot to handle complex, multi-step tasks by testing outcomes in its mind before committing to physical actions. It bridges the gap between understanding a scene and physically interacting with it, creating a unified brain for real-world problem solving. The result is a robot that thinks visually, making it far more capable and adaptable in dynamic environments.

🆔 @hackernewsgithubprojects

📦 noizai/worldsonus

WorldSonus: Watch, Listen, and Change the Sound

WorldSonus turns silent video clips into synchronized stereo soundtracks that you can actually steer. Instead of just guessing what noises belong on screen, this tool watches the visual motion and listens to your written instructions to generate realistic audio in tiny one-tenth-of-a-second chunks. The coolest part is that you can change the description while the video plays, so the sound updates in real time to match new actions. It works by linking what you see with what you hear, making sure a splash or a train horn lands exactly when it happens on screen.

🆔 @hackernewsgithubprojects

📦 something11235/dsh-video-skin

dsh-video-skin

A live video can now play behind your chat window and roll as a sound-on boot animation in the same interface. The project called dsh-video-skin turns a static desktop into a moving surface, letting you drop in your own clips for the background and startup splash. It solves the problem of a boring, flat screen by layering video directly under the UI without any external dependencies. The coolest part is the built-in slider that adjusts transparency in real time, so you can balance visibility and readability instantly. It is a zero-dependency tool for people who want their app to feel like a piece of media rather than just a window.

🆔 @hackernewsgithubprojects

📦 betelgeuse-ai/betelgeuse

Betelgeuse: Running Massive AI on a Budget GPU

Run frontier-scale AI models directly on a modest consumer graphics card. Betelgeuse is a specialized fork of a popular local inference engine designed specifically to handle massive mixture-of-experts architectures without needing a data center. Instead of forcing the entire model into limited video memory, it keeps the core structure on the graphics card and intelligently pulls in specific expert modules from your fast local drive only when they are actually needed. This clever paging technique lets a budget setup handle computations that would normally overwhelm it, turning a standard gaming rig into a capable local inference station.

🆔 @hackernewsgithubprojects

📦 batterworks/hatchdoor

Hatchdoor: Give AI Access To Your Notes

Hatchdoor lets your AI agent read and edit your personal notes directly from a self-hosted server. It turns a folder of plain text files into a smart interface where a computer can browse, search, and update your documents. This is useful because your AI assistant can actually change your notes without you having to copy and paste everything manually. It acts like a bridge between your local files and the artificial intelligence you use daily. You keep your data on your own computer, giving you full control over your information. This project is great for people who want their AI to interact with their personal knowledge base securely.

🆔 @hackernewsgithubprojects

📦 battermanz/hatchdoor

Hatchdoor: Let your AI edit your notes

Hatchdoor is the self-hosted app that lets your AI agents read, search, and edit your personal notes. Most note-taking apps are locked down, but this tool opens up your Obsidian-style vault so any AI assistant you connect to it can actually do work. You get a fast web interface to browse and manage files yourself, plus a backend connection that lets your favorite chatbot find that one specific paragraph you forgot about. It is built for people who want their digital library to be active, not just a static archive.

🆔 @hackernewsgithubprojects

📦 signalfx/obstudio

Obstudio: See Your Code's Health

Obstudio turns your local development setup into a live health monitor for your software. It captures the actual activity of your code as it runs, turning raw data into clear visual dashboards you can inspect right in your browser. The standout feature is its ability to connect directly with AI coding assistants. It gives these tools the specific context they need to automatically find missing performance data and fix it for you. Instead of guessing why an app feels slow, you get precise, real-time visibility into exactly what is happening. This makes debugging faster and less stressful for developers.

🆔 @hackernewsgithubprojects

📦 gentel-lab/fold2reason

Fold2Reason: The AI That Learns To Think Like A Protein

Fold2Reason is the post-training technique that teaches AI to reason better by studying how proteins fold. Researchers at Gentel Lab took a standard language model and trained it to predict the physical shapes of proteins, using a special spatial workspace and a frozen geometry decoder. The surprising result? This biological training made the model significantly better at general logic, math, and science questions it had never seen before. It seems that learning to understand the three-dimensional structure of life actually forces the AI to develop deeper, more robust reasoning skills.

🆔 @hackernewsgithubprojects

📦 csr1010/wappy-kit

Wappy: Your WhatsApp AI Agent

A working artificial intelligence agent can live on your own computer and talk back on your existing WhatsApp number, without sending a single byte of your data to a cloud server. The open-source project Wappy solves the privacy gap in personal assistants by letting developers run a smart, memory-capable bot locally that connects directly to the public WhatsApp business API. This means you build a private digital helper that remembers context and retrieves your own notes entirely on your own hardware. You bring your own model key and phone number, keeping total control over your data. No vendor lock-in and no forced subscriptions.

🆔 @hackernewsgithubprojects