📦 0xbakeer/qwen3.8-27b-4-bit-on-a-single-dgx-spark

Run Qwen 27B Fast on One Chip

A twenty-seven billion parameter model hits seventy-five tokens per second on a single DGX Spark chip. This repository, named Qwen 3.8 27B 4 bit on a single DGX Spark, is a measured proof that you do not need a server farm for this level of performance. It tests how squeezing the model into four-bit memory actually changes speed. The big surprise is that the speed boost only matters when serving one user at a time. Once you add a crowd, the advantage vanishes entirely. It also shows which settings break the model and which ones are safe.

🆔 @hackernewsgithubprojects

📦 tejasq/basically-ai-harness

basically-ai-harness: The Scaffolding That Makes AI Useful

basically-ai-harness is the minimal TypeScript tool that proves an AI model needs a structural cage to work. Built for a talk on harness engineering, it separates the raw model from the scaffolding that manages tools and memory. The key insight here is that guardrails catch structural errors, but only a verification step catches a confident lie. By giving one model a real browser and another a fake one, the repo shows that the harness is the moat. The model is rented, but the environment is yours.

🆔 @hackernewsgithubprojects

📦 franksuperg/electrochemical-battery-model-atlas

The Battery Model Atlas Explained

Sixteen out of nineteen complex battery simulations actually run without breaking. That is the core promise of the Electrochemical Battery Model Atlas. It is not a simulator itself, but a verified map of existing tools. Instead of guessing which codebase works, you get a checklist that proves which models run on your computer. It saves you weeks of debugging broken dependencies by showing you exactly what runs and what does not. You get a clear path from theory to working code. This is the fastest way to trust your simulation results.

🆔 @hackernewsgithubprojects

📦 nagara214/ssr-merge

SSR-Merge: Merging AI Styles Without Retraining

SSR-Merge takes separate AI models trained on different subjects and fuses them into a single, coherent image generator without any extra training. Usually, combining these specific style filters requires expensive retraining or results in messy, conflicting outputs. This tool uses a clever routing method to blend the signals, letting you merge a cat filter and a dog filter into one unified model that draws both subjects clearly in the same picture. It runs a quick calibration on a single graphics card and saves the result as a standard file you can load anywhere. You get the flexibility of multiple specialized models in one lightweight package, all done in minutes.

🆔 @hackernewsgithubprojects

📦 zz1358m/latent-memory-master

Latent Memory Master: Squeeze a Document into One Token

A single token can hold the meaning of a whole document. That is the core idea behind Latent Memory Master. This research project solves a major problem for small language models: they usually forget long texts because their memory is too short. Instead of just cutting text, this system compiles long documents into compact latent memory slots. It forces a temporary compressor to squeeze all important facts into a few hidden vectors. Then, it throws away the compressor entirely. The frozen base model only sees these tiny memory slots to answer questions. This allows small, resource-constrained models to handle large contexts without massive computational costs.

🆔 @hackernewsgithubprojects

📦 wangyongpeng90/cuckoo-code

Cuckoo Code: The AI Agent That Costs Zero Tokens

Cuckoo Code turns your free web browser AI into a hands-on coding assistant that can actually do things. It runs the official web version of DeepSeek or Claude inside a local window and quietly listens to the conversation. When the AI suggests a command or file change, Cuckoo Code catches it, shows you what is about to happen, and runs it right on your computer. The coolest part? You pay absolutely nothing in extra token fees because it just borrows your existing free account. You get a real agent that can read your files, run scripts, and fix bugs without you typing a single line of code.

🆔 @hackernewsgithubprojects

📦 cluvexstudio/aether

Aether: The Unblocker

Aether automatically hunts down working routes on heavily censored networks so you can browse freely. It hides your traffic inside standard web requests, making it look like ordinary HTTPS traffic that deep packet inspection tools cannot easily block. The program finds these paths on its own, tests them to ensure they actually pass data, and then opens a local proxy for your apps to use. You point your browser at this proxy, and the software handles the rest, connecting through Cloudflare’s infrastructure to bypass regional restrictions. It is a practical tool for anyone facing digital walls, offering a way to restore normal internet access without complex setup.

🆔 @hackernewsgithubprojects

📦 unstablebuild/rune

Rune: The IDE Built for Speed

Rune is the GPU-rendered IDE that finally respects your keyboard. It bundles your code editor, terminal, and debugging tools into one lightning-fast, composable window. The real kicker? It ships an AI agent as a simple extension, so your machine can help you code without messing up your manual workflow. It is built for people who live in the terminal and want everything to feel instant. If you are ready to trade slow, clunky tools for something that actually keeps up with your thoughts, this is the one to watch.

🆔 @hackernewsgithubprojects

📦 jtydhr88/screenwriting-skills

Screenwriting Skills: Your AI Co-Writer

Screenwriting Skills is the collection of agent plugins that turns your AI assistant into a seasoned screenwriting mentor. It distills the wisdom from forty-seven craft books and twenty-three volumes of published scripts into twenty-six specific skills you can install once and use in both Claude Code and OpenAI Codex. The standout feature is its deep dive into Chinese opera and television production, offering rare, granular guidance on traditional opera structures and mainland China series formats that you will not find in standard Hollywood writing guides.

🆔 @hackernewsgithubprojects

📦 alchaincyf/huashu-mac-use

Huashu Mac Use: The Invisible Agent

Huashu Mac Use is the agent skill that lets AI control your Mac’s native apps without stealing your focus. Most automation tools force you to stop working whenever a script runs, but this one operates in the background so you can keep typing while it fills forms and clicks buttons for you. The coolest part is how it proves its work by taking screenshots of the exact window it touched, even if that window is hidden behind others. It reads your screen without interrupting you and only borrows focus when absolutely necessary, leaving a clear trail of every action taken.

🆔 @hackernewsgithubprojects

📦 speedstu/cuda-for-amd-windows

Cuda For AMD On Windows

A real neural network trainer completed a full training cycle on a standard AMD graphics card running Windows. This setup proves that software written exclusively for Nvidia chips can actually learn on AMD hardware. The project bridges the gap by intercepting Nvidia commands and translating them into instructions your AMD chip understands. This means you can run existing artificial intelligence tools without needing to rewrite your code or buy a new graphics card. It is currently validated on one specific model, but the path is open for others. If you are tired of waiting for official support, this working stack is your proof that it is possible today.

📰 https://news.ycombinator.com/item?id=49684356

🆔 @hackernewsgithubprojects

📦 viceverse-cz/serein

Serein: The Discord Client That Actually Works

Swap out the bloatware for a client that actually respects your hardware. Serein is a lightweight Discord desktop app built entirely in Rust, ditching the heavy web engine for a native, lightning-fast experience. While the official client hogs over a gigabyte of RAM, Serein sips just a fraction of that, running in a single process with near-zero idle CPU. You still get everything you need, including voice calls, screen sharing, and a slick interface, without the background drain. It is a genuine performance win for your PC, proving that lighter is faster and way better for your system.

🆔 @hackernewsgithubprojects

📦 arikchakma/gpu-time

gpu-time

A tiny neural network lives entirely inside your browser, turning messy human text into precise calendar dates without sending your data to a single server. Called gpu time, this tool solves the frustrating problem of reading times like 'every other Tuesday' or 'Saturday to Sunday' without a massive cloud connection. It runs right on your computer's graphics card, making it fast and private. You type in natural language, and it gives you exact start times and repeating rules instantly. It is a clever way to understand dates locally, keeping your schedule yours. If you want to parse time without the cloud overhead, this small model is the simple, smart answer.

🆔 @hackernewsgithubprojects

📦 letsfg/letsfg

The Travel Tool That Cuts Out the Middleman

LetsFG is the travel search engine that gives your AI agent direct access to real, un-inflated flight and hotel prices. It solves the problem of cookie tracking and surge pricing by reading directly from airlines and booking sites, so the same search always returns the same honest price. You can connect it to Claude or Cursor via a simple server setup, or just use the command line. It scans hundreds of airlines and major sites to find the cheapest deal, then holds the fare on your card while a booking agent buys the ticket for you.

🆔 @hackernewsgithubprojects

📦 clickhouse/walshadow

walshadow: Sub-Second Postgres to ClickHouse Replication

Send your Postgres data to ClickHouse in under a second. That is the whole point of walshadow. Most data pipelines lag behind your database by minutes or hours, but this tool grabs the raw write-ahead log and mirrors your rows into ClickHouse almost instantly. It even handles the messy stuff like schema changes and initial loads without breaking a sweat. If you are tired of slow analytics, this is the fix. It turns your primary database into a real-time source for your analytics engine. No more waiting for batch jobs to finish. Just pure, fast data flow. It is experimental, but the speed is real.

📰 https://news.ycombinator.com/item?id=49699321

🆔 @hackernewsgithubprojects

📦 overment/limen

Limen: The File-Based AI Team Harness

Limen turns a single chat command into a fully autonomous software development team without installing a new platform. It works by spawning real AI agents as separate processes inside isolated Git branches, letting them code, test, and review work while you handle the merge. The coolest part is that the entire workflow is built from plain files and Git, meaning there are no hidden databases or opaque servers, just your code. You tell a coordinator what to build, and it hands the job to a worker in its own sandbox. When the task is done, you simply review the branch like any other pull request.

📰 https://news.ycombinator.com/item?id=49700087

🆔 @hackernewsgithubprojects

📦 j92580498-max/pockethle

PocketHLE: Run 2000s Mobile Games Today

PocketHLE runs Windows Mobile games from the early two-thousands directly on your modern phone, laptop, or desktop. It skips faking the entire operating system and instead runs the actual game code while simulating just the specific system calls it needs to work. That means you can boot titles like Call of Duty two on a Linux machine or an Android phone. It is an open-source project built in Rust, and it is still early stage, but it is a clever way to keep those old handheld games alive without a dedicated retro device. If you have old mobile games sitting in a drawer, this is the tool to watch.

🆔 @hackernewsgithubprojects

📦 shadcn-ui/lint

Lint: The Design System Linter Built for AI

Coding agents break design systems because they lack context. Lint fixes this by turning your specific design rules into actionable error messages that tell an AI exactly which component variant or theme token to use instead of a raw value. When an agent tries to hardcode padding, the tool explains why that violates your button component and suggests the correct size prop, effectively translating your design intent into instructions the machine can follow. This makes automated code generation respect your brand guidelines without rewriting your components. It is not just a checker; it is a guide that keeps AI-built interfaces consistent with your established look and feel.

🆔 @hackernewsgithubprojects

📦 hypit-ai/hypit

How Hypit Clones Viral Videos

Hypit is the AI video pipeline that turns one viral clip into a hundred unique variations with a single command. It works by analyzing your source video to automatically generate fresh scripts, swap out the on-screen talent, and create entirely new B-roll footage. The result is a massive batch of tailored versions ready for upload, removing the heavy lifting of manual editing. You simply point it at a trending style and let the system do the heavy creative work.

🆔 @hackernewsgithubprojects

📦 mutonby/openshorts

OpenShorts

OpenShorts stacks two people vertically in a single frame so captions land on the gap between them instead of covering a face. This open source tool watches your long form video, finds the best moments, and cuts them into vertical shorts that are ready to post. It uses artificial intelligence to crop the frame tightly around the speaker so nobody gets hidden, then adds subtitles and a hook text to grab attention. You can run this entirely on your own computer for free, keeping your data private while turning podcasts or webinars into content for TikTok and Instagram.

📰 https://news.ycombinator.com/item?id=47300951

🆔 @hackernewsgithubprojects

📦 apeireth/apeireth

Apeireth: The AI OS That Actually Thinks

Apeireth is the operating system that finally treats your AI like a brain, not just a chatbot. Most AI tools forget everything the moment you close the tab, but Apeireth uses a continuous memory system to remember facts and learn from every interaction. It acts like a digital mind that keeps getting smarter the more you use it. It’s written in Rust, so it’s fast and safe, but the real magic is that it actually remembers. If you want an AI that grows with you instead of starting from scratch every time, this is the one to watch.

🆔 @hackernewsgithubprojects