📦 ashalliants/ninfer-3090

Ninfer-3090: The AI Engine for One Card

Ninfer-3090 is the inference engine that makes a massive twenty-seven billion parameter AI run fast on a single consumer graphics card. Most people think you need a server farm of expensive chips to run this class of model, but this project squeezes every drop of performance out of a single RTX 3090. It uses specialized math techniques to process the model’s data much faster than standard tools, letting you get answers at real-time speeds without breaking the bank. It’s a masterclass in making heavy software run on normal hardware, proving that you don’t need a data center to run the latest tech.

🆔 @hackernewsgithubprojects

📦 restsend/opus-rs

Opus-RS: Audio Compression Without the Heavy Lifting

The entire Opus audio codec is now running on a microcontroller with zero heap memory. Opus-RS is a pure Rust implementation of the standard audio compression format used by Zoom and Discord. The original C version needs a standard operating system to run, which locks you out of embedded devices and web browsers. This project solves that by removing all dynamic memory allocation, allowing the encoder to run on bare metal hardware with no global allocator. It performs just as well as the C version, proving you don't need a heavy runtime to compress high-quality audio.

🆔 @hackernewsgithubprojects

📦 dioxuslabs/rustd

rustd: The Fast Rust Builder

Rustd swaps out your standard build command for a significantly faster replacement that compiles your Rust code much quicker without changing how you write it. If you spend hours waiting for your project to compile, this tool tackles that exact pain point by acting as a drop-in upgrade to your existing setup. The interesting part is the speed: it reworks the way the compiler processes your code, turning long wait times into short pauses. For anyone building large Rust applications, this means iterating on ideas without the usual lag. It is a practical tool for developers who want their machines to work as hard as they do.

🆔 @hackernewsgithubprojects

📦 comfy-org/comfy-openjutsu

Comfy Openjutsu: Swap Video Characters

Comfy Openjutsu is the video tool that swaps the actor in a clip with a character from a single image. You upload a short video segment and a character picture, then the app replaces the person in the footage with that character while keeping the original background, movements, and sound exactly as they were. It is a standalone web app built on a specific Comfy workflow that you can clone and run locally. The best part is how realistic it feels; the character matches the lighting and scene of your original video seamlessly.

🆔 @hackernewsgithubprojects

📦 stewdioai/blank-page

Blank Page: The Skill That Fixes Your Vague Prompts

Blank Page solves the frustrating moment when you have a vague idea for a video clip but no clear visual direction. Instead of asking you to describe every detail, this skill takes your one line of text and turns it into a fully detailed prompt. It does this by automatically finding five reference images to lock in the style and three motion examples to define the movement. It then writes out a numbered sequence of beats for your clip, including a creative twist like a zoom or a reveal. The result is a rich, specific prompt that gets you a much better first draft than a generic request.

🆔 @hackernewsgithubprojects

📦 pch/omafeed

omafeed: Your Offline RSS Command Center

omafeed is the native Linux reader that treats your news feed like a scriptable database. It keeps your subscriptions and articles entirely on your computer, but the real surprise is the built-in command line that lets you query, search, and star articles from a terminal without opening the window. You can pipe this data into scripts or give it to an AI agent to build custom daily digests. It is a privacy-focused tool that works offline, uses no cloud sync, and offers a clean three-pane view for human reading.

🆔 @hackernewsgithubprojects

📦 thusi-lab/prismwam

PrismWAM: The Robot Brain That Splits Sight From Motion

PrismWAM is the robot model that finally separates what a scene looks like from how it moves. Most AI systems get confused when the lighting changes, but PrismWAM solves this by splitting its thinking into two distinct parts. It uses a simple color snapshot to understand the goal, then relies on detailed depth maps to plan the actual physical movement. This clever trick allows it to guide real robots with impressive accuracy, even without years of extra training. If you are building autonomous machines, this approach to vision is a major breakthrough.

🆔 @hackernewsgithubprojects

📦 twehrbein/exmr

ExMR: Full Body Motion Capture from One Camera

ExMR is the motion capture tool that finally nails down hand and foot details from a single camera. Usually, figuring out exactly how someone moves in three dimensions requires expensive, marker-covered suits. This project solves that by taking an ordinary video and reconstructing a complete, ground-level skeleton. It goes beyond just tracking the main body; it specifically refines the tricky movements of your fingers and toes. That level of detail is what animators and robot engineers need to make their characters feel real. You can export this data directly into game engines or animation software. It turns your phone into a professional motion capture studio.

🆔 @hackernewsgithubprojects

📦 einsia/embodiedevo

EmbodiedEvo: The Robot That Learns By Doing

Watch a coding agent figure out how to fold clothes, stack bowls, or play tic-tac-toe by actually moving a robot in a simulation. This project connects an AI coder to a physics engine, letting it watch what happens, rewrite its own control code, and try again until it works. The coolest part is how it remembers. After practicing on easy tasks, it saves reusable tricks and written notes about what went wrong. Then, when it faces a brand-new challenge, it starts fresh with those notes in hand, acting like a seasoned intern who learned from past mistakes.

🆔 @hackernewsgithubprojects

📦 hazyresearch/ncm

Next-Concept Modeling: The Fast Text Generator

Next-Concept Modeling is the text generation framework that writes whole sentences in parallel instead of word by word. It solves the slow speed problem of standard language models by using a two-stage system. First, it sketches a rough version of your text, then it refines that sketch into a detailed paragraph in just a few steps. This means it generates text up to thirty-one times faster than other methods while keeping the same quality. It’s like watching a rough draft turn into a polished essay in seconds.

🆔 @hackernewsgithubprojects

📦 tyrichards/omarchy-tray

omarchy-tray: The System Tray That Swallows Your Bar

omarchy-tray is the system tray plugin that turns your entire desktop bar into a hidden drawer. If you use Omarchy, you know how cluttered the bar gets with clocks, workspaces, and weather widgets. This project solves that mess by letting you drag any bar widget directly into the tray, tucking it away until you hover the chevron to slide out the drawer. It is a strict upgrade over the stock tray because it keeps all your standard features like pinning and app menus while adding this new capability.

🆔 @hackernewsgithubprojects

📦 tool-bunker/code-buster

Code Buster: Your Code's GPS

Code Buster tells an A I coding agent exactly where its changes broke the rules, all without sending your code to the cloud. It acts as a local guardrail that maps every dependency and logic cycle in your project before the agent writes a single line. Instead of guessing, it gives the machine a deterministic map of dead code and architecture violations, so it can fix errors while it still has context. It’s like giving a new intern the entire building’s blueprints before they start moving furniture. No more hallucinated imports. Just grounded, offline verification that keeps your AI honest and your codebase clean.

🆔 @hackernewsgithubprojects

📦 igtoth/tracedraw

TraceDraw: The Vector Editor Rebuilt in Rust

TraceDraw is the open-source vector editor that finally brings full-featured illustration to your browser and desktop without bloat. This project is a clean-room rewrite of a classic drawing suite, built entirely in Rust to run natively on Windows, macOS, Linux, and even in your web browser via WebAssembly. It solves the problem of heavy, proprietary software by offering a lightweight, native experience that still handles complex tasks like live effects, page layout, and print-ready PDFs. The most surprising part is that it natively opens and writes Corel Draw files, meaning you can swap between old projects and this new tool without losing your work.

🆔 @hackernewsgithubprojects

📦 vsmutok/ytscrape

ytscrape

ytscrape lets you pull raw data from YouTube without the usual hurdles. It skips the need for an official API key and avoids the slow, clunky process of running a fake browser in the background. Instead, it talks directly to the same internal data routes that the website itself uses. This means you can grab search results, video details, and even the full text of every comment and caption in seconds. It is a clean, lightweight way to collect information for research or building tools, giving you actual data without the heavy setup or constant worry about hitting usage limits.

🆔 @hackernewsgithubprojects

📦 hao-ai-lab/ora

ORA: Letting Open Models Drive Robots

An open-weight language model just picked up a robot arm and started solving physical puzzles without a single line of robot-specific training code. This project, called ORA, acts as a bridge that lets standard AI models operate physical robots on the fly. Instead of needing expensive, custom-trained programs for every new task, the system simply gives the model tools to see the world and move its limbs. It then figures out how to stack blocks or pour liquid entirely on its own. This is huge because it proves that general intelligence can handle real-world physical actions without years of specialized data collection.

🆔 @hackernewsgithubprojects

📦 alibaba/last-secure-code-benchmark

Do AI Coders Break Security?

Watch top AI coding agents fail to write safe code when left alone. The Last Secure Code Benchmark tests if these models can rebuild a project without accidentally leaving a hole in the door. Here is the wild part. Even the best models only pass every single security check on a quarter of the tasks. This happens because the tests never tell the agent to think about safety. They just ask for working code, then secretly try to hack the result. It turns out that building something that runs correctly is not the same as building it safely.

🆔 @hackernewsgithubprojects

📦 dmitryryabkov/local-ai-mac

Run a 128K AI on a MacBook

Turn a spare Mac into a private AI server that handles massive context windows without ever sending your data to the cloud. Local AI Mac is a practical guide that shows you exactly how to squeeze real power out of Apple’s unified memory. The coolest part? You can run a thirty billion parameter model with the same one hundred twenty eight token context as top cloud services, all locally. This means your code and chats stay on your machine, and you stop paying per token. If you have thirty two gigabytes of RAM or more, you can actually do this.

📰 https://news.ycombinator.com/item?id=47754997

🆔 @hackernewsgithubprojects

📦 arnaudguiovanna/tutor-mcp

Tutor MCP: The AI Tutor That Actually Grades You

Your AI assistant just started refusing to let you guess your way through a lesson. Tutor MCP is an open source tool that plugs into any large language model to turn it into a strict, scientific tutor. It doesn't just chat; it maps your knowledge into a graph and uses fifty years of cognitive science to track exactly what you know. If you guess wrong, it freezes your progress and forces you to practice until you actually understand the material. It’s like having a patient professor who remembers your mistakes and schedules reviews so you never forget them.

📰 https://news.ycombinator.com/item?id=49744316

🆔 @hackernewsgithubprojects

📦 plotarmordev/drama-tensorfold

Drama Tensorfold: 3.7x Faster Voice on DGX Spark

Run Fish Audio's Drama three locally on a DGX Spark and get your first audio chunk in just fourteen hundredths of a second. That is the single coolest thing in this repo: a patched engine that makes text-to-speech respond almost instantly while staying under forty-eight seconds for long-form audio. It cuts the time to generate a typical eight-second clip by nearly half, turning a sluggish local server into a snappy, responsive tool for real-time apps. The repo includes the exact engine patches, build recipes, and benchmarks so you can verify these speed gains on your own hardware.

🆔 @hackernewsgithubprojects

📦 punchisoft/punchi-dock-remastered

Punchi Dock Remastered: Audio Visualizer for Linux

Punchi Dock Remastered turns your Linux dock into a live audio visualizer by reading your system's sound output through PipeWire. The dock icons themselves become the spectrum analyzer, displaying real-time sound waves in six different styles. It’s not just a pretty effect; the whole dock is a native, customizable workspace for apps, folders, and media controls. It blends seamlessly into KDE Plasma, letting you drag, drop, and organize your desktop with fluid animations. If you want your Linux desktop to feel as alive as your music, this is the tool to try. It’s a genuine step forward for desktop aesthetics.

🆔 @hackernewsgithubprojects