📦 tonyd2wild/glm-5.2-quanttrio-200k-4x-dgx-spark--36tok-s

glm-5.2-quanttrio-200k-4x-dgx-spark--36tok-s

The specialized cluster recipe that serving teams are using to run the massive GLM-5.2 model at full quality without deleting any of its two hundred fifty-six experts. Instead of surgically hacking the model down to fit on your hardware, this setup orchestrates four NVIDIA GB10 nodes to host the unpruned model at a massive two hundred thousand token context. It squeezes out up to thirty-six tokens per second using smart trickery like built-in speculative decoding and custom Triton kernels. It even includes an indexer patch to prevent multi-user crashes. If you want to deploy full-scale reasoning models across high-end hardware, grab this recipe, spin up your cluster, and start serving.

🆔 @hackernewsgithubprojects

📦 xang1234/stock-screener

Stock Screener

This multi-market stock scanner lets you run institutional-grade technical and fundamental scans across twelve global markets simultaneously, including the US, India, and Hong Kong, without paying for expensive data feeds. The platform orchestrates complex trading strategies like William O'Neil's CANSLIM and Mark Minervini's trend templates, combining eighty technical filters with an AI chatbot to research trending themes from live feeds. It solves the massive headache of managing scattered market data and high API costs by pulling information into a single, Docker-deployed system backed by PostgreSQL and Redis. Anyone looking to level up their trading workflow gets a self-hosted, industrial-strength screening pipeline that runs entirely on local hardware.

🆔 @hackernewsgithubprojects

📦 drsexo/frosty

Frosty

You can now reclaim your Android phone's battery life by freezing power-hungry background services with a single tap. Frosty is an advanced Magisk, KernelSU, and APatch module designed specifically to stop background battery drain and rein in persistent Google services. Through its clean web interface, it lets you disable unnecessary telemetry, automate screen-off power saving, and apply deep sleep optimizations without breaking your favorite features. It even auto-tunes system memory and block storage performance based on your device's actual hardware. It is the ultimate tool for anyone wanting complete, smart control over their phone's standby power.

🆔 @hackernewsgithubprojects

📦 open-compass/agentcompass

AgentCompass

Benchmark and evaluate your AI agents under realistic conditions using this unified testing suite. It runs parallel tests across over twenty major benchmarks to systematically grade how well your language models handle complex, multi-step operations like writing code, executing terminal commands, browsing the web, and navigating visual interfaces. By spinning up secure, isolated environments using Docker or Modal, it lets you safely watch your agents attempt real-world software engineering tasks and deep research assignments. You get back clear metrics and step-by-step trajectory graphs showing exactly where an agent succeeded or got stuck. Grab the repository to start stress-testing your agents today.

🆔 @hackernewsgithubprojects

📦 harahan/meanflownft

MeanFlowNFT

Train image and video generators to produce high-quality, aesthetic results in only four steps. This project introduces a clever way to apply reinforcement learning directly to fast, velocity-based AI generators without destroying their speed or efficiency. Usually, fine-tuning these models with human feedback rewards requires slow, complex calculations during training. This framework solves the problem by optimizing the model's trajectory in a simulated space while keeping the actual generator lightweight and fast. In practice, this means you can generate stunning media using Stable Diffusion and video models with incredible detail, beating older, slower methods. Give this repository a try to see how efficient AI generation can really be.

🆔 @hackernewsgithubprojects

📦 katipally/openlive

OpenLive

OpenLive is the open-source voice and vision layer that finally gives your favorite AI agents eyes and ears right on your own machine. Instead of renting an expensive, locked-in cloud pipeline, this clever tool runs your entire voice loop locally using WebGPU. It handles the listening, the speaking, and even screen sharing privately, letting you connect to almost any brain you want. You can voice-drive coding agents like Claude Code, answer permissions aloud, and even clone your own voice in seconds. It is the ultimate way to talk to your AI without paying extra audio fees.

🆔 @hackernewsgithubprojects

📦 siw00-lim/tango

tango

You can now edit 3D models with incredible precision without drawing a single manual mask. A clever new framework called tango makes 3D editing entirely training-free by steering the generative process in what mathematicians call tangent space. Instead of applying a clumsy global update that ruins the parts of your model you want to keep, it automatically targets only the areas that need to change. By amplifying guidance for tokens in the edit zone and softening it everywhere else, it keeps the original object's identity perfectly intact. It is a brilliant, math-driven shortcut to painless, localized 3D modifications.

🆔 @hackernewsgithubprojects

📦 pku-sec-lab/jetson-pi-edge

jetson-pi-edge

Run massive vision-language-action robot control models directly on low-power onboard hardware in real time. Running advanced artificial intelligence on a physical robot usually means dealing with painful lag that ruins smooth movement. This repository solves that by packing smart speedups like graph reuse and GPU-resident buffers into a lightweight runtime based on llama.cpp. It squeezes incredible performance out of compact devices like the NVIDIA Jetson, slashing latency by more than half so your robot can react instantly to visual feedback. Just spin up its persistent background server, feed in your camera feeds and robot states, and watch it generate lightning-fast physical actions.

🆔 @hackernewsgithubprojects

📦 internrobotics/real

Real

The robotics framework that finally bridges the gap between simulated training and real-world physical deployment. Real uses raw camera feeds and the Model Context Protocol to guide mobile robots through complex tasks like navigating rooms, finding objects, and opening doors. Instead of guessing, these robots can actually communicate with humans to clear up confusing instructions. It solves the classic simulation-to-reality bottleneck by matching virtual environments perfectly with physical hardware, achieving an incredible seventy-eight percent success rate in actual physical tests. If you want to see how robots are learning to interact with our open world, check out the Real repository.

🆔 @hackernewsgithubprojects

📦 michaelshimeles/adam

Adam

Adam is the clever port that takes Vercel's durable AI agent runtime, called Eve, and runs the entire execution engine directly inside Convex. Usually, running these complex, long-lived agents requires managing a heavy Node server, persistent databases, and message queues. Adam completely eliminates the server by compiling your agent loop and running it directly in Convex's serverless environment, keeping the workflow state, schedules, live streams, and UI together in one unified place. It is a fantastic project for developers because it simplifies deployment and ensures that even hours of complex AI steps run reliably without server maintenance. Check out Adam to see how serverless AI architecture is becoming incredibly simple.

🆔 @hackernewsgithubprojects

📦 qualtyco/api-doctor

api-doctor

The code linter that prevents AI-generated code from breaking your external integrations. While artificial intelligence is amazing at writing quick software drafts, it often hallucinates incorrect API structures, leaves out critical web verification steps, or hardcodes sensitive keys. To solve this, api-doctor uses deterministic structure-based rules rather than unpredictable AI models to scan your files. It instantly flags broken configurations for popular platforms like Supabase, Firebase, and OpenAI before they crash in production. Running this simple tool in your terminal ensures your third-party integrations stay secure, reliable, and functional without relying on another AI to double-check the first one.

📰 https://news.ycombinator.com/item?id=48762860

🆔 @hackernewsgithubprojects

📦 amap-cvlab/abot-agentos

abot-agentos

Coordinate smart robots with a dedicated cognitive operating system that sits directly above physical hardware. While modern robots are excellent at seeing and moving, they struggle with long-term memory and learning from their mistakes. abot-agentos fixes this by serving as a thoughtful mental layer that manages reasoning, tracks physical skills, and verifies actions in real time. It features a persistent multi-modal memory that transforms visual observations, conversations, and spatial layouts into a structured knowledge graph. Best of all, a built-in self-evolution loop automatically diagnoses task failures and adapts, ensuring the robot constantly learns and gets smarter over time.

🆔 @hackernewsgithubprojects

📦 nowork-studio/toprank

toprank

The open source tool toprank gives your AI agent direct control over your marketing channels, letting it audit Google Ads, Meta Ads, and SEO performance, find wasted spend, and automatically execute the fixes right from your terminal. It connects your AI assistant directly to live data to spot things like negative keyword opportunities or broken landing pages. Instead of just showing you a static dashboard, it actually writes meta tags, adjusts bids, pauses underperforming keywords, and structures schema markup. It is perfect for developers who want automated, real-time marketing optimization without leaving their terminal or dealing with complex manual setups.

🆔 @hackernewsgithubprojects

📦 kurtgokhan/tegaki

Tegaki

Tegaki is the web utility that instantly turns any standard font into realistic, self-drawing handwriting animations. Instead of forcing you to manually map out path data or wrestle with bulky native software, this lightweight tool handles the entire process directly in the browser or terminal. It automatically takes your chosen font, reduces it to a clean digital skeleton, and animates it stroke by stroke with natural, hand-drawn timing. It works right out of the box with major frameworks like React, Vue, and Svelte, and it even lets you generate animated SVGs from your command line. Simply pick a font and watch your text write itself.

🆔 @hackernewsgithubprojects

📦 anthropics/financial-services-plugins

financial-services-plugins

financial-services-plugins is the blueprint collection that finally gives AI agents the precise skills needed for complex finance workflows. Instead of relying on generic prompts, this repository provides pre-built, domain-specific agents and plugins designed for intensive tasks like investment banking, equity research, and private equity. It contains specialized skills that let Claude draft pitch decks, build financial models directly in Excel, audit general ledgers, and parse onboarding documents. You can run these tools directly as user-facing plugins or deploy them as managed agents behind your own custom system. Grab these templates to build your own intelligent financial analyst today.

🆔 @hackernewsgithubprojects

📦 ryancodrai/py-turboquant

py-turboquant

If you want to run massive vector search but are tired of your server's memory crying, py-turboquant has a brilliant solution. This tool implements Google's TurboQuant algorithm in Rust with Python bindings, and it can compress a thirty-one gigabyte corpus of vector data down to just four gigabytes while actually searching it faster than standard FAISS. Instead of requiring a separate, slow training phase to learn your data's distribution, it uses a clever random rotation trick to make any incoming vectors predictable, meaning you can add new embeddings instantly. It is perfect for building incredibly fast, highly compressed local search applications right on your machine.

📰 https://news.ycombinator.com/item?id=47562135

🆔 @hackernewsgithubprojects

📦 khcrysalis/impactor

Impactor

Impactor is the cross-platform sideloading tool that finally lets you install custom apps onto your Apple devices without needing a Mac. Built entirely in Rust, it works like a mini version of Xcode right on your computer, whether you run Windows, Linux, or macOS. It registers your device, grabs the proper certificates using your Apple ID, signs the apps locally, and handles the installation automatically. You can even inject tweaks, change app names, or generate pairing files to keep everything running smoothly. If you want a clean, simple way to manage your mobile apps on your own terms, this project is a total game changer.

🆔 @hackernewsgithubprojects

📦 ulsklyc/oikos

Oikos

Oikos is the self-hosted family planner that finally puts you in complete control of your household data without subscription fees. Instead of throwing your private routines onto corporate servers, this clever tool lets you run your own family dashboard on any home setup. It bundles everything from chore charts, calendars, and meal planners to shopping lists and budget trackers into one offline-friendly web app. The absolute best part is its modular design, meaning you only toggle on the features your family actually uses. It is a brilliant way to keep your household completely organized, private, and independent.

🆔 @hackernewsgithubprojects

📦 vudovn/antigravity-kit

Antigravity Kit

Antigravity-kit is the local developer tool that finally turns your AI code editor into a fully coordinated, multi-agent engineering team. Instead of letting a single AI chat get overwhelmed by complex tasks, this project injects a pre-configured configuration folder directly into your workspace. It comes packed with twenty specialized developer personas, forty-seven modular technical skills, and thirteen interactive slash commands. The system automatically routes your requests to the right domain expert and uses persistent memory to remember your project guidelines across sessions. Download it to upgrade your coding assistant into a structured, highly efficient software development department.

🆔 @hackernewsgithubprojects

📦 affaan-m/everything-claude-code

everything-claude-code

You can now transform standard AI coding assistants into a highly organized, domain-specific developer fleet. The everything-claude-code repository is a comprehensive optimization framework designed to supercharge tools like Claude Code and Cursor. It provides pre-built system prompts, specific agent personalities, and automated commands that guide AI models through complex coding, testing, and security tasks. Instead of getting generic AI suggestions, you gain specialized virtual teammates, from rust reviewers to database architects, that follow strict project rules. Grab this setup to instantly make your AI coding agents smarter, safer, and far more practical for everyday software development.

🆔 @hackernewsgithubprojects