📦 azampatti/qwen3.8-flash-next-int4-fast

Run a 125B Model on One GPU: Qwen3.8-Flash-Next

Run a 125 billion parameter AI model on a single desktop chip. Qwen3.8-Flash-Next does this by shrinking the model to four-bit precision and cutting the number of active experts in half. It runs at roughly seventy tokens per second on a single DGX Spark, which is surprisingly fast for such a large brain. The project includes scripts to download the weights and start the server in minutes. You get a local, OpenAI-compatible API that supports tool calling and reasoning. It is not just a static file; it is a complete recipe for local inference. The takeaway is simple: you do not need a data center to run frontier-scale models anymore.

🆔 @hackernewsgithubprojects

📦 jouaaaaaa/mcd-self-butler

McButler: The McDonald's Butler That Actually Thinks

McButler is the McDonald's butler that actually thinks. It uses the official McDonald's API to stop you from guessing at calories and prices. The coolest part? It creates a credit score for your burger habits. It looks at your order history and treats your loyalty like a bank loan. It checks your spending rhythm and taste variety to see if you are a stable customer or a chaotic snack. It is a clever way to make a fast food order feel like a financial audit. Next time you order, let the AI judge your burger discipline.

🆔 @hackernewsgithubprojects

📦 surendrapandar/minilytics

Minilytics: Know exactly which tweet made you money

A web analytics tool that tells you exactly which single tweet made you money is a rare find. Most dashboards show you page views, but Minilytics connects every visitor to the actual sale they made. It tracks where people came from, whether that was a social post, a search engine, or even an AI assistant. Then it links that journey to a completed purchase. This is huge for solo founders who need to know what actually works. You host it for free on Cloudflare, keeping your data private. Stop guessing which marketing channel pays off. Minilytics gives you the exact revenue source behind every dollar you earn.

🆔 @hackernewsgithubprojects

📦 ayushjain1144/trackeverything

TrackEverything: How It Tracks Every Point in 3D

TrackEverything is the tracking system that finally keeps every single point in a scene moving smoothly through long videos. It solves a major headache for computer vision: usually, when a camera moves or objects shift, tracking algorithms lose sight of things or get confused by overlapping images. This project fixes that by building a persistent three-dimensional memory of the space. It removes duplicate points so the computer isn't wasting energy tracking the same pixel twice. This allows it to follow objects across thousands of frames without getting lost.

🆔 @hackernewsgithubprojects

📦 dsojevic/juxtopposed-vlc

Juxtopposed VLC: The Redesign That Actually Works

A massive redesign of the world's most famous media player is finally playable. Juxtopposed VLC takes a stunning visual concept that used to live only in design files and turns it into working software. This is not just a skin; it is a full adaptation of the interface for the real player, with responsive layouts and interactive controls. It keeps the engine that plays every file you throw at it, but wraps it in a fresh, modern look built with the Inter font. If you have ever wondered what a sleek, professional interface for your favorite open source tool would look like, this fork is your answer.

🆔 @hackernewsgithubprojects

📦 fermionresearch/phonon

Phonon: Open Speech Recognition

Phonon turns raw audio files into clean, accurate text right on your own hardware, completely free of any cloud dependency. This open source engine runs directly on your laptop or desktop, whether you are using a Mac, a Windows PC, or a Linux box, so your sensitive meetings never leave your machine. It handles everything from quick command line scripts to live microphone dictation, and even serves as a local server that mimics standard speech APIs. Because it is fully open, you can inspect the code and trust the process without worrying about third party data handling.

🆔 @hackernewsgithubprojects

📦 lyramakesmusic/neuron-explorer

Neuron Explorer: See What AI Neurons Actually Write

Turn up a single brain cell inside a large language model and watch the whole thing start talking like an alien. That is exactly what Neuron Explorer does for the Qwen three point five four B model. Instead of guessing what a specific part of the network is for, this tool physically pushes that part harder and records every word the model generates under the pressure. It builds a three dimensional map so you can see which neurons make the model talk about space, and which ones just make it stutter. It is a way to open the black box and see the actual thoughts emerging from the math.

📰 https://news.ycombinator.com/item?id=50233138

🆔 @hackernewsgithubprojects

📦 designfuchen/word-atlas

Word Atlas: The Illustrated IELTS Companion

Word Atlas is the study tool that turns every word into a picture. Most vocabulary apps give you a list of text and expect you to memorize it, but this project assigns a unique original illustration to each of the seven thousand and two hundred and ninety-eight core words for the IELTS exam. You see a car left in the snow to understand the word abandon, or steam rising from a volcano to learn its name. This visual link makes the meaning stick in your memory much better than a simple definition.

🆔 @hackernewsgithubprojects

📦 nordstjernen-web/northstar-web-browser

Northstar Web Browser: Written From Scratch In C

Northstar Web Browser is a complete web engine written entirely in C, built from scratch rather than modified from existing giants. It handles modern web pages with its own JavaScript interpreter and video decoders, all inside a single, lightweight process. The project proves that a fully functional, secure browser can exist without the massive complexity of standard industry engines. This code is small enough for one person to read and audit end to end. It is a clear reminder that the web runs on simple, transparent code, not just on the biggest names in the industry.

🆔 @hackernewsgithubprojects

📦 0xbakeer/qwen38-flash-next-spark

Run 180B Model on One Desk

Run a 180 billion parameter model on a single desktop box. That is the whole trick. You can serve a massive AI model with a full 262,000 token context on just one NVIDIA DGX Spark unit. The secret is that half of the model is actually just a giant lookup table. It does not do any math. It just gets read. So you keep that huge table sitting on your hard drive instead of in memory. This frees up space for the actual thinking parts to fit inside the computer. It is not just a hack. It actually works well. You get real speed.

🆔 @hackernewsgithubprojects

📦 nlzy/turing-unlock

turing-unlock

Stop letting your graphics card baby its matrix math. Turing unlock is a tiny Linux tool that flips a hidden switch to let the half-precision cores run at full throttle. NVIDIA artificially caps these chips, but this script pokes directly at the hardware to remove that ceiling. The result? Your card suddenly processes data nearly twice as fast as before. It is a bit risky because the extra speed generates serious heat, but for anyone running local models, it is a massive, free performance boost. Just be careful. This is not a safe little tweak. It is a raw override that can push your hardware to its limits.

🆔 @hackernewsgithubprojects

📦 huggingface/carbon

Carbon Genomic Foundation Model

Carbon lets you treat DNA sequences just like text, allowing a standard language model to predict the next genetic code without any fine tuning. Trained on trillions of base pairs from the Carbon Pretraining Corpus, this family of models learns the hidden grammar of life, distinguishing healthy genes from harmful mutations by simply reading the sequence. This solves a major hurdle in genomics where testing variants usually requires expensive wet lab work, making it possible to screen for disease risks directly from digital data. The flagship three billion parameter model matches or beats larger rivals, proving that a smaller, focused model can deliver powerful biological insights.

🆔 @hackernewsgithubprojects

📦 hazymorning/origami

Origami: The Dashboard Your Phone Actually Needs

Your home assistant phone screen is usually a messy grid of buttons that feels less like a tool and more like a spreadsheet. Origami fixes this by delivering a complete, ready-made dashboard where the interface itself is the product, not just a color theme. The most striking part is how it handles pop-ups for specific devices like your fan or thermostat. When you tap a tile, the interface opens a dedicated window showing the exact speed, oscillation, or temperature, turning a simple tap into a proper remote control experience. It even loads its own font directly from your server so the look is perfectly consistent without waiting on outside connections.

🆔 @hackernewsgithubprojects

📦 microsoft/agensh

Agensh: 1024 AI Agents Working Together

Coordinate a thousand AI agents to solve complex coding tasks without a single manager. Microsoft’s Agensh creates a self-organized team where each agent claims specific jobs, shares discoveries, and verifies peer results through shared workspaces. This decentralized approach proves that larger groups significantly outperform single individuals on difficult programming benchmarks. The system naturally develops specialized roles and coordination workflows as the team expands, showing how collective intelligence scales effectively. This project offers a clear blueprint for building autonomous, multi-agent systems that work together seamlessly. If you are exploring how to scale AI capabilities, this is the essential model to study for your next project.

📰 https://news.ycombinator.com/item?id=50223759

🆔 @hackernewsgithubprojects

📦 unclebob/ub-mode

ub-mode

Force your AI coding assistant to actually think before it types a single line. ub-mode is a skill file you drop into tools like Claude or Cursor that stops the agent from just guessing and makes it follow a strict engineering ritual. It forces the AI to write a failing test first, hunt down the root cause of a bug instead of just patching over it, and use specific tools to prove the code isn’t broken. It essentially turns a chatty chatbot into a disciplined engineer who checks its own work.

🆔 @hackernewsgithubprojects

📦 clarkcheekykent/cheekyfoveateddlss

Cheeky Foveated DLSS Boosts VR Frame Rates

Cheeky Foveated DLSS is the graphics tool that makes your virtual reality games run faster without ruining the picture. The project works by realizing that your eyes only focus sharply on the very center of your vision, while the edges are blurry anyway. It uses this natural blind spot to tell your graphics card to process the crisp details only in that small center circle, skipping the heavy math for the blurry parts you cannot see. This simple trick saves a huge amount of computer power, resulting in massive frame rate gains even in demanding titles.

🆔 @hackernewsgithubprojects

📦 hffgzktbcc-blip/ounce-core

Ounce-Core: The Rust Switch 2 Emulator

Ounce Core runs unaltered Nintendo Switch 2 game binaries directly on your computer by writing a complete virtual console from scratch in Rust. Instead of relying on old code, this project builds a fresh emulation engine that understands the new Switch 2 hardware, including its updated processor and graphics system. It handles everything from loading game cartridges to managing sound and graphics, creating a fully functional software replacement for the physical hardware. The most impressive part is that it executes real, unmodified game files without needing to tweak them first, proving it can handle the latest console technology.

🆔 @hackernewsgithubprojects

📦 jordanhubbard/daemonarchy

Daemonarchy: The FreeBSD Desktop

Daemonarchy is the bootable FreeBSD image that brings the polished Omarchy desktop to the BSD world without sacrificing system safety. It solves the friction of installing a curated Linux-style environment on a Unix system by packaging the entire setup into a single USB stick that handles everything from disk partitioning to driver installation. The standout feature is the boot environment system, which creates a complete, bootable copy of your entire system every time you update. If an update breaks something, you simply restart and select the previous snapshot to roll back instantly. It turns the complex operating system into something manageable for daily use.

🆔 @hackernewsgithubprojects

📦 colbyjack1134/spaceengineersvr

Play Space Engineers in Real VR?

Playing a complex space sandbox game through a virtual reality headset without a dedicated VR interface is usually a mess of awkward button presses and invisible targets. Space Engineers VR fixes this by turning your motion controllers into actual hands you can point, grab, and poke with. The most surprising part is how it handles the cockpit. Instead of relying on a standard keyboard, the plugin maps your physical controller sticks directly to the ship's flight controls, letting you fly and drive with your wrists. It also includes a smart system that lets other mods recognize when you are in VR so their buttons and menus stay visible and interactive.

🆔 @hackernewsgithubprojects

📦 catcatuncle/openworkbuddy

OpenWorkBuddy: The AI That Actually Makes Your Files

OpenWorkBuddy is the local-first AI agent that finally hands you real, editable files instead of just a chat log. You know the frustration of asking a bot for a report and getting a wall of text you have to manually copy and format? This tool fixes that. You tell it what you need, and it actually generates the PowerPoint, Word doc, or Excel sheet on your own computer. It uses your own AI keys, so your data stays local and private, which is a huge win for privacy. The best part? You can add new skills just by dropping in a simple text file.

🆔 @hackernewsgithubprojects