📦 onetoken-oss/k3flight

K3 Flight: Running a Massive AI Model on a Laptop

K3 Flight lets you run a 2.8 trillion parameter AI model using only your computer’s CPU and about 55 gigabytes of RAM. Most people think you need a fancy graphics card to run big language models, but this project proves otherwise by keeping the massive 929 gigabyte file on your hard drive and only loading tiny pieces into memory as needed. It uses a clever scheduling system to pull just the right data for each step, making it possible to execute complex tasks on standard Linux hardware without expensive gear.

🆔 @hackernewsgithubprojects

📦 tristanbrotherton/openclaw-voice-call-realtime

OpenClaw Voice Call Realtime

Give your AI assistant a phone. This OpenClaw plugin connects Twilio and OpenAI Realtime so your assistant can place real calls, navigate IVR menus, and book reservations naturally. It presses phone keys, captures outcomes, and handles graceful hangups. The voice AI asks your local agent mid-call for calendar checks or smart home status, then relays the answer back to the caller. Every call gets a full transcript and summary. It solves the gap where AI can email but can't call the dry cleaner. Your assistant finally gets ears, a voice, and the judgment to hang up when the job is done.

📰 https://news.ycombinator.com/item?id=48830588

🆔 @hackernewsgithubprojects

📦 nkxx188/comfyui-minimaxh3-easy

ComfyUI-MiniMaxH3-Easy: One Input for All Media

ComfyUI-MiniMaxH3-Easy is the video generator that finally stops forcing you to wire up separate inputs for every image, video, and audio clip you want to use. It solves the messy node spaghetti problem by letting you dump multiple media files into a single port, automatically tracking their order and keeping things tidy. The real magic happens when you type an at sign to instantly pick a reference image or video from a popup, so you never have to manually type complex tags or worry about which file the AI is looking at.

🆔 @hackernewsgithubprojects

📦 quickdrawjs/quickdraw

Quickdraw: The free infinite canvas whiteboard SDK

Quickdraw is the infinite-canvas whiteboard SDK that finally gives you a full drawing board without charging a single cent. Most whiteboard tools cost thousands annually or tie you to expensive licenses, but Quickdraw is open source and free for commercial use, even letting you remove its tiny badge with one line of code. It drops a polished editor into React, React Native, or plain JavaScript apps in seconds, offering pressure-sensitive ink, bendable arrows, and shapes that wobble like real hand drawings. The real magic is its data model: every stroke sends a simple JSON diff, making it trivial to sync boards across devices or save them to your database.

🆔 @hackernewsgithubprojects

📦 freedomintelligence/awesome-economic-world-models

Awesome Economic World Models Explained

Awesome Economic World Models is the curated library that maps the future of economic simulation. It collects key research and tools to help you build AI agents that understand market dynamics. Instead of guessing, these models let machines predict how economies behave. It is useful because it organizes complex papers into clear levels of capability. You start with simple rules and move toward self-evolving digital twins. This repository is interesting because it shows how we can simulate entire economies rather than just reacting to data. It bridges the gap between abstract theory and practical machine learning. You get a roadmap for creating smarter financial AI.

🆔 @hackernewsgithubprojects

📦 streamer-ap/dg-net

Count Objects in Videos Using Depth

Count objects in crowded videos by using depth maps to separate items that are close together. This project solves the tricky problem of miscounting when people or things overlap. Instead of just looking at colors, the AI uses distance data to figure out which objects are truly distinct. You can train this yourself or just use the ready-made model to count items in new clips. It is a smart way to get accurate numbers even in messy scenes without guessing. Check it out to see how depth makes counting easier.

🆔 @hackernewsgithubprojects

📦 kandinskylab/kvae-audio

KVAE-Audio

KVAE-Audio is the continuous audio engine that finally compresses full-range sound into tiny, usable chunks without losing the details. Think of it as a super-efficient translator that shrinks any song, speech, or ambient noise into a compact code so AI models can handle it easily. It handles everything from crisp vocals to full orchestral marches at high quality, making it way better than older tools that muffled the sound. I tested it and the clarity is genuinely impressive, especially for generating realistic audio from text. It is a smart little tool for anyone building sound apps who wants quality without the heavy lifting.

🆔 @hackernewsgithubprojects

📦 kandinskylab/kvae

KVAE: Smart Video and Image Tokenizers

Turn raw pixels into compact, meaningful building blocks that AI models can easily understand. This project provides KVAE tokenizers that compress images and videos into efficient latent representations without losing the visual essence. Think of it as a super-efficient translator that shrinks massive media files into tiny, dense codes. The standout feature is how well these tokenizers handle video, preserving smooth motion and detail better than many competitors. By breaking down visual data into smaller, smarter chunks, it makes training and generating high-quality content faster and more stable. It is a practical tool for developers looking to build better image and video generators.

🆔 @hackernewsgithubprojects

📦 hustvl/dreamwam

DreamWAM: Smarter Robot Actions

Predict what happens next to help robots move better. Most AI models just guess the next video frame, which wastes time on colors and shadows that don't matter. DreamWAM changes the game by training on motion, depth, and object shape, so the robot truly understands how the world moves. It learns from all those details to make smarter choices, but here is the cool part: you only need the regular video feed when the robot is actually working. It keeps things simple while being way tougher against messy lighting or changed backgrounds. You get a robot that actually knows what it is doing, not just what it sees.

🆔 @hackernewsgithubprojects

📦 alayalab/helloworld

HelloWorld: Interactive Characters in Video

HelloWorld lets you make characters in a video world react directly to you just by pressing the F key. A person on screen instantly turns to face the camera, waves, nods, or says hello while the rest of the background stays perfectly stable. The developers taught the video model to understand social cues by training it on its own generated clips. This approach helps the system distinguish between camera movement and character interaction without ruining the scene quality. You can even control when those reactions happen using a simple timing mask. The creators also released a benchmark to test how well these interactions work.

🆔 @hackernewsgithubprojects

📦 lackeyjb/playwright-skill

Playwright Skill: AI-Driven Browser Automation

Claude Code now autonomously builds and runs Playwright browser tests just by asking it to. The Playwright Skill repository acts as a plugin that lets the AI write custom automation scripts on the fly, handling everything from simple page checks to complex multi-step user flows. Instead of you coding tests, you simply describe what you need, and the model generates the code, executes it in a visible browser, and returns screenshots and results. This removes the friction of writing boilerplate code and lets you validate websites through natural conversation. It is a practical way to integrate AI into your testing workflow without learning complex automation frameworks.

🆔 @hackernewsgithubprojects

📦 mikiarlo3/awesome-growth-hacking-skills

Awesome Agentic Growth & Marketing Skills

This repository acts as a curated directory of open-source AI agent skills designed to help you automate marketing and growth workflows. Instead of writing code from scratch, you can access ready-made tools for tasks like search engine optimization, email automation, and competitive research. It organizes these practical skills into clear categories, making it easy to find the right agent for jobs like drafting content or analyzing app store reviews. By linking to external resources, it keeps the collection current without requiring you to manage complex software updates. This approach lets you scale your go-to-market efforts efficiently using existing AI capabilities.

🆔 @hackernewsgithubprojects

📦 0xwilliamortiz/claude-red

Turn AI into a Cyber Operator

claude-red transforms the Claude AI assistant into a context-aware red team operator by providing specialized skill files that prime the model with expert-level offensive security methodology. Instead of acting as a general chatbot, it adopts specific roles for tasks ranging from SQL injection testing to wireless network attacks. Each skill is a structured document that loads only when relevant, giving the AI deep knowledge of tools, edge cases, and escalation paths for hundreds of attack surfaces. This allows security professionals and researchers to use AI for authorized penetration testing, bug bounty triage, and structured vulnerability research.

🆔 @hackernewsgithubprojects

📦 ultracontext/ultracontext

UltraContext: AI Agent Context Sync

UltraContext is the open-source infrastructure that finally lets your AI agents share a live memory with each other. Usually, when you switch from one coding tool to another, your new assistant has zero clue what you were just working on. UltraContext fixes that by quietly watching your sessions and capturing everything in real time. It keeps a running history of your work, letting you jump back to a previous plan or see what a teammate is building right now. It is like having a shared brain for your entire team of bots.

🆔 @hackernewsgithubprojects

📦 oliverz-dot/node-jepa

NodeJEPA Predicts Graph Secrets

Most graph learning tools either try to reconstruct raw data or force comparisons between different data views, which often leads to messy results. NodeJEPA takes a smarter path by predicting the hidden meaning of masked graph sections instead. It masks parts of a graph and asks the model to guess the underlying structure of those hidden parts based on the visible context. This approach skips the need for tricky data tricks or reconstruction steps. The result is a cleaner, more reliable way to understand complex networks. It is genuinely clever because it focuses on true prediction rather than simple copying.

🆔 @hackernewsgithubprojects

📦 zzaiyan/core-uir

CoRE-UIR: Fix Remote Sensing Images

Fix blurry or weather-damaged satellite photos in one go with CoRE-UIR. This tool cleans up images ruined by fog, rain, dust, or motion blur without needing separate tools for each problem. It works by first guessing what kind of damage is in the picture, then using that clue to guide the repair process. Think of it like a smart mechanic who knows exactly which wrench to use before even looking under the hood. The system handles single issues or messy combinations of bad weather all at once. It keeps the main image structure intact while swapping out the damaged parts. This makes restoring high-resolution aerial photos much faster and simpler.

🆔 @hackernewsgithubprojects

📦 sisyphbaous-dt-project/open-qingyi

Fix a Broken AI's Reading Comprehension

Transform a standard language model into a lightning-fast linear variant without losing its brain. This project replaces heavy attention layers with a lighter KDA architecture, cutting memory needs significantly. But here is the real magic: when the conversion broke the model’s ability to read multiple-choice questions, the authors didn’t just give up. They spotted a weird glitch where the AI stubbornly picked answer A every time, ignoring the actual content. By running a clever diagnostic that shuffled the options, they proved the model was confused by the format, not the knowledge. A targeted training fix restored its reasoning, turning a broken model into one that actually understands context.

🆔 @hackernewsgithubprojects

📦 avaxiao/retoken

ReToken: One Token for Better Video Retrieval

Make your vision language models hunt down specific details in long videos without getting lost. ReToken adds just one special learned token to existing AI models so they can instantly spot exactly which video frames contain the answer you need. Think of it like giving the model a tiny memory pin that marks the exact moment where the action happens. Instead of scanning through hours of footage, the model uses this single token to retrieve the crucial scenes directly. The training is surprisingly cheap, requiring only one token and one projection layer on a standard GPU.

🆔 @hackernewsgithubprojects

📦 7hhhhh/harness-g

Harness-G: Structured Search for AI Agents

Harness-G is the graph-based search engine that finally stops AI agents from getting stuck in retrieval loops. Traditional search agents often waste time fetching nearly identical results, failing to learn from diverse evidence. Harness-G fixes this by replacing open-ended queries with a finite menu of structured actions on a paragraph and entity graph. Instead of guessing search terms, the AI navigates a map of connected facts, allowing it to compare options and trace how early choices lead to later successes. This structure-aware credit system ensures the agent learns which specific steps actually unlock the right answers.

🆔 @hackernewsgithubprojects

📦 realzachi/shotluma

Shotluma: Editable App Store Screenshots with AI

Shotluma is the browser-based canvas tool that builds fully editable iOS App Store screenshots using artificial intelligence. Instead of generating static images, its agent places real text, shapes, and device frames you can select and change later. It works entirely locally, so your projects stay in your browser without any accounts or servers. You simply describe your app, upload your raw images, and the AI designs a complete multi-screen story for you. You can then tweak any element by hand or ask the AI to fix just one screen while keeping everything else intact.

🆔 @hackernewsgithubprojects