📦 tonyd2wild/deepseek-v4-flash-dspark-1m-nvfp4-kv-2x-dgx-spark

deepseek-v4-flash-dspark-1m-nvfp4-kv-2x-dgx-spark

Run massive AI models with a massive one-million-token context window on just two nodes. This project provides a validated recipe to serve the DeepSeek V4 Flash model with speculative decoding across a pair of high-performance servers. By utilizing a highly compressed attention cache format, it fits nearly two million total tokens into active memory. It specifically solves the problem of model degradation, like endless loops or random text drift, when running high-concurrency agent workflows. Now, developers can easily handle multiple long-context sessions simultaneously without crashing their hardware or sacrificing speed. Give this setup a spin to scale your local AI agents.

🆔 @hackernewsgithubprojects

📦 balukosuri/andrej-karpathy-s-autoresearch-as-a-universal-skill

Andrej Karpathy's Autoresearch as a Universal Skill

Drop this self-improving prompt optimizer into any code repository, and it automatically scans your codebase, suggests what to optimize, defines clear evaluation metrics, and runs an autonomous loop that makes your prompts better over time while you sleep. Inspired by Andrej Karpathy's method of letting an AI agent autonomously run and evaluate code experiments, this universal tool brings that exact evolutionary process to prompt engineering. Instead of manually tweaking prompts based on vibes, the system defines strict binary yes-or-no criteria to grade performance, runs continuous tests, keeps what works, and throws away what doesn't. It is the ultimate hands-off way to build battle-tested prompts for any project.

🆔 @hackernewsgithubprojects

📦 vasilyevdm/ai-agent-handbook

ai-agent-handbook

The ai-agent-handbook is the master engineering reference that finally pulls back the curtain on how thirty of the biggest AI agent frameworks actually work under the hood. Instead of surface-level tutorials, this project did a deep dive into real production codebases like LangGraph and Claude Code to figure out why agents fail and how to fix them. The coolest takeaway is how they handle context rot, showing that agent performance drops when the context window is only a quarter full because bloated tools eat up all the memory. It is a goldmine of real architectural patterns for building smarter, leaner AI systems.

📰 https://news.ycombinator.com/item?id=47461861

🆔 @hackernewsgithubprojects

📦 timoncool/higgsaudio-studio

Higgs Audio Studio

Generate expressive speech and fully direct your own audiobooks and podcasts locally with a clever desktop application. higgsaudio-studio packages a powerful voice generator and a local brain that acts as your personal text director, automatically inserting emotion and sound effect tags to make the audio feel incredibly alive. You can clone voices, host multi-speaker dialogues with perfect volume leveling, and keep character voices consistent throughout long audiobooks. The best part is that everything runs completely offline on your own machine with no subscription fees. Grab your microphone and start creating professional-grade audiobooks in your room today.

🆔 @hackernewsgithubprojects

📦 tachikomared/character-animation-creator-skill

character-animation-creator-skill

character-animation-creator-skill is the game development tool that turns text prompts or reference images into fully animated, pixel-art spritesheets. If you are building an indie game, this project makes it easy to generate matching walk, idle, and attack animations from a single concept picture. Instead of drawing every frame by hand, it coordinates with AI image generators to create horizontal frame strips, cleans up messy background colors, snaps the pixel grid, and packages everything into clean sheets. It automatically handles multiple resolutions and eight-direction movements so your characters are ready to drop straight into your favorite game engine.

🆔 @hackernewsgithubprojects

📦 serenakeyitan/awesome-notebooklm-prompts

awesome-notebooklm-prompts

awesome-notebooklm-prompts is the curated directory that finally rescues your AI presentations from looking like boring corporate templates. This collection gathers highly creative, field-tested prompts sourced from design communities and power users to completely transform how you generate slides. By feeding these precise design specifications into NotebookLM, you can instantly turn raw text, research, or notes into beautiful, highly stylized decks. Instead of generic bullets, you get distinct visual themes like sharp-edged minimalism, retro-developer grids, bold newspaper layouts, and vibrant street styles. It is the ultimate shortcut to making automated slides look incredibly intentional and professionally designed.

🆔 @hackernewsgithubprojects

📦 karaage0703/xangi

xangi

xangi is the chat assistant hub that finally bridges your favorite messaging platforms directly to powerful local and cloud agent backends. Instead of being locked into a single interface, you can talk to your assistant through Discord, Slack, Telegram, LINE, or even a simple local browser window. Under the hood, it routes your requests to smart developer tools like Claude Code, Cursor, or local offline models. It even broadcasts live updates over a local stream so physical desktop smart toys can react to its thoughts. It is the ultimate way to self-host a fully automated, connected digital brain.

🆔 @hackernewsgithubprojects

📦 keivalya/mini-vla

Mini-VLA

Mini-VLA is an educational robot brain that makes learning robot control accessible by packing a complete vision-language-action model into about one hundred and fifty lines of core code. Instead of dealing with massive, complex AI systems, you can see exactly how a robot combines camera images, text instructions, and its own physical state to decide how to move. It uses a clever setup to mix this sensory information and guide the robot's physical actions step-by-step. It is a brilliant, bite-sized playground for anyone wanting to understand how modern robots actually think and translate words into real physical motion.

🆔 @hackernewsgithubprojects

📦 turbostarter/loading-ui

Loading UI

Loading UI is the component library that finally makes building beautiful, polished loading states for your web apps incredibly simple. It is a curated, open-source collection of modern spinners, loaders, and animations designed specifically for web developers. Instead of wrestling with complex CSS or heavy libraries, you can browse through a variety of sleek, pre-designed loading animations to instantly match your app's style. It solves the headache of creating custom transitions from scratch, giving your users a delightful experience while they wait. You can easily find a design you love, grab the code, and drop it straight into your project.

🆔 @hackernewsgithubprojects

📦 maddog241/mvinverse

mvinverse

Reconstruct detailed 3D scene geometry and materials from a sequence of images in just a few seconds. A new open-source project called mvinverse makes this possible by skipping the slow, traditional optimization process entirely. Instead, it uses a smart attention mechanism to look across multiple views and directly predict consistent material and surface properties. If you have ever tried to turn photos into 3D assets, you know how easily reflections and lighting can ruin the results. By processing everything in a single, fast forward pass, this tool keeps colors and textures perfectly aligned across every angle. It is an incredibly fast way to turn regular images into accurate 3D materials.

🆔 @hackernewsgithubprojects

📦 pacifio/ui

ui

ui is the design language that teaches AI coding agents how to build beautiful, dense, AMOLED-black interfaces. Instead of manually copying and pasting CSS components, this project operates as a specialized skill for tools like Claude Code. It provides a structured set of design rules and token references that guide your favorite AI assistant to generate consistent, ultra-compact layouts in plain HTML, Tailwind, or React. It solves the messy, inconsistent UI code that agents typically write by giving them a clear aesthetic framework to follow. You get pixel-perfect, dark-mode designs built directly by your AI companion.

🆔 @hackernewsgithubprojects

📦 b1u3b01t/design-playground

Design Playground

You can now visually redesign your React components using local AI, right inside your Next.js project. A developer-focused tool called design-playground installs nested inside your existing app folder, leaving your main package configuration completely untouched. Once started, it automatically scans your workspace to find your React components and loads them onto an infinite, interactive canvas. From there, you can drag your components around, write a quick prompt, and watch Cursor or Claude Code generate layout and style variations side-by-side. When you find a design you like, you simply copy the generated code directly back into your project.

📰 https://news.ycombinator.com/item?id=48750398

🆔 @hackernewsgithubprojects

📦 elinglijiaoqiao/presentation-ppt-maker

Presentation PPT Maker

Presentation-ppt-maker is a structured slide generator that turns a pile of academic papers into a complete, ready-to-present slide deck. Instead of relying on a single AI prompt that might hallucinate facts or create chaotic layouts, this tool coordinates four specialized agents to extract figures, organize claims, draft a slide-by-slide narrative, and programmatically generate the final presentation file. Its standout strength is a mandatory visual verification gate that actually opens image files to ensure figures match the slide text before placing them, automatically splitting crowded slides if a diagram is too small to read. You get perfectly cited, beautifully organized slides with complete speaker notes ready for your next talk.

🆔 @hackernewsgithubprojects

📦 ramziddin/solid-skills

Solid Skills

Transform junior-level code into senior-engineer quality software using an AI agent skill that enforces rigorous coding standards. This repository provides a structured set of rules and reference documents that guide AI coding assistants to follow test-driven development, write tiny methods under ten lines, and completely avoid raw primitive types for domain concepts. Instead of allowing messy code, it prompts the AI to wrap everything from emails to money in dedicated value objects, catch code smells automatically, and apply design patterns naturally. It is the perfect tool for developers who want automated, clean, and highly maintainable architecture right out of the box.

🆔 @hackernewsgithubprojects

📦 jtydhr88/comfytv

ComfyTV

ComfyTV is the canvas-based app that finally turns ComfyUI into an intuitive, node-by-node creation playground. Instead of running a massive, tangled workflow that recalculates from scratch every time you tweak a setting, this clever setup lets you run each stage individually. You can generate an image, edit it, paint in details, and convert it to video in a structured pipeline where downstream steps simply use snapshots of previous outputs. It gives you rich built-in tools like a 3D camera, panorama viewer, and mask painter without ever leaving the canvas. Grab this tool to finally make your local generation workflow feel fast, organized, and truly interactive.

🆔 @hackernewsgithubprojects

📦 sermuns/ratatuefi

ratatuefi

Build sleek terminal dashboards that boot directly on your bare metal computer without any operating system running underneath. A clever project called ratatuefi bridges the gap between the popular Ratatui terminal UI library and raw UEFI firmware. It gives you a minimal, zero-bloat backend written in Rust, allowing you to draw styled terminal interfaces straight from your motherboard's bootloader. By mapping interface colors and cursor positions directly to UEFI console outputs, you can run gorgeous, lightweight diagnostic tools or custom boot menus. It is the perfect weekend rabbit hole if you want to make your bare-metal booting sequence look incredibly cool.

🆔 @hackernewsgithubprojects

📦 cedarconnor/spag4d

spag4d

You can now transform a single flat three-sixty-degree panoramic photo into a fully explorable three-D Gaussian Splat scene in as little as two seconds. A clever tool called spag4d uses advanced depth-estimation models to calculate spherical geometry and project flat pixels directly into three-D space. It handles everything from simple depth projection to complex multi-face stitching using Apple's SHARP model. If your scene has structural holes where the camera couldn't see, the software can even generate synthetic walkthrough videos to intelligently patch and fill those gaps. It is an incredibly fast way to bring flat panoramic memories to life.

🆔 @hackernewsgithubprojects

📦 arteemg/autoswarm

AutoSwarm

Autoswarm is the open-source tool that finally makes your local artificial intelligence models smarter over time just by using them. It acts as a clever proxy sitting in front of your offline models, silently logging your chats and running a background review to distill valuable lessons into a custom skillbook. The next time you ask a question, it automatically injects those lessons back into the prompt, and it even prunes away old strategies that did not work out. It is a brilliant way to make a basic offline setup continuously learn from its own mistakes without burning a ton of extra computing power.

🆔 @hackernewsgithubprojects

📦 jd-opensource/joyai-echo

JoyAI-Echo

JoyAI-Echo can generate synchronized, multi-shot video and audio sequences up to five minutes long while keeping character faces and voices perfectly consistent across every scene. It solves the massive headache of identity drift and audio mismatch in AI-generated films using a smart cross-modal memory bank that remembers what characters look and sound like. Even better, it runs over seven times faster than older systems, allowing you to quickly create editable stories from a single text prompt. If you want to experiment with long-form, coherent AI filmmaking, you can run JoyAI-Echo locally or use its official ComfyUI node integration.

🆔 @hackernewsgithubprojects

📦 robzilla1738/supergoal

Supergoal

Supergoal lets you run massive, complex programming tasks entirely on autopilot using Claude Code or Codex. Instead of baby-sitting an AI coding agent step-by-step, you run a single command describing what you want to build. The tool scans your codebase, anticipates technical risks, and splits the work into logical phases. It then gives you one final command to paste, which launches an autonomous execution loop. If a step fails, the system automatically attempts a self-healing recovery using a three-strike retry protocol. Once finished, it runs a rigorous final audit to double-check that every single change is fully functional.

🆔 @hackernewsgithubprojects