📦 joshuaswarren/omarchy-mlx

omarchy-mlx: Apple Silicon Linux AI

omarchy-mlx is the Linux driver that finally lets Apple silicon run machine learning without macOS. It takes the M-L-X framework, which normally only works on Apple's proprietary metal graphics interface, and swaps it out to run on standard Vulkan drivers instead. This means you can run local chat models and speech recognition entirely on your Mac hardware while staying on Linux, with no cloud accounts and no proprietary lock-in. It is essentially a bridge that keeps the tensors on the GPU, giving you a private, local AI stack on hardware you already own.

🆔 @hackernewsgithubprojects

📦 heqiangdiao-wq/ai-cross-border-ecommerce-agent-toolkit

The AI Toolkit for Cross-Border Ecommerce

Take a veteran’s best practices and turn them into a repeatable recipe for selling online. This project solves a specific problem in international ecommerce. It turns the messy, gut-feel work of experienced sellers into seven clear AI workflows. Instead of guessing, you get exact prompts and templates for everything from market research to ad optimization. It even includes scripts to crunch real data so the AI isn’t just guessing. The coolest part? It’s a closed loop. It doesn’t just generate a product listing and leave you hanging. It analyzes your sales and customer reviews, then tells you exactly what to fix to improve your results.

🆔 @hackernewsgithubprojects

📦 localai-org/localvqe

LocalVQE: The Tiny AI That Cleans Your Calls

LocalVQE acts as a tiny, intelligent filter that strips the annoying echo from your speakerphone calls without needing a graphics card. It works by listening to both your microphone and the exact sound playing out of your speakers, allowing it to mathematically subtract that feedback loop in real time. The surprising part is how incredibly light this is. A specialized version runs on a single Raspberry Pi core, processing audio twenty-one times faster than real time. This means you get studio-quality, echo-free voice on any basic laptop or even a low-power device. No cloud servers, no heavy hardware, just a small file that makes your audio instantly clearer.

🆔 @hackernewsgithubprojects

📦 orange-3dv-team/mocha

Swap Video Characters Instantly with MoCha

MoCha swaps the person in any video with a new character using just one image and a single frame mask. It works without needing complex body tracking or depth maps, which makes it far more reliable than older tools that break when characters move or overlap. The system learns from a large set of synthetic training videos to handle tricky lighting and motion smoothly. If you have ever wanted to change who is in a clip without re-acting the scene, this tool does it in a single pass. It is a powerful way to test out new characters or fix a bad take right from your desktop.

🆔 @hackernewsgithubprojects

📦 piggybackbayonet/miauto

MiAuto: Your Phone's Car Brain

MiAuto is the Android app that turns your phone into a full car infotainment system without a dongle. It runs on your phone and acts as the receiver side of the ICCOA CarLink protocol, letting you wirelessly project your phone's screen and audio onto your car's head unit. You get navigation, media, and controls right on your car's dashboard, no cable, no extra hardware, and no need to root your phone. It's basically your phone becoming the car's brain, handling the connection so your phone's apps appear on the car's screen.

🆔 @hackernewsgithubprojects

📦 sola1039/hast-sec

hast-sec

Make a digital avatar move naturally by predicting its next motion directly from its own recent history. Hast Sec is a research project that uses a small neural network to analyze the angular velocity of a character's bones. Instead of manually animating every frame, the model predicts how the skeleton will continue moving based on the previous four frames of data. It handles everything from subtle head turns to full body shifts, ensuring the motion feels fluid and continuous rather than robotic. This is interesting because it solves the problem of stiff, choppy animations by learning the natural physics of movement.

🆔 @hackernewsgithubprojects

📦 black-forest-labs/flux2

FLUX.2: Sub-Second Image Generation

Generate and edit images in under a second on a standard gaming computer. Black Forest Labs just released FLUX.2, a family of open-weight models that run locally on your own hardware. The surprise here is the klein variant, which creates high-quality visuals or edits existing photos almost instantly, fitting entirely within eight gigabytes of memory. This means you can build interactive creative tools without expensive cloud servers or waiting for results. If you are looking for a fast, local way to experiment with visual generation, this is the one to try.

🆔 @hackernewsgithubprojects

📦 scrapecreators/social-media-research-skills

Social Media Research Skills: Let Your AI Agent Find Viral Secrets

Hand your AI agent a single prompt and it hunts through public social data to find the exact comments where people are confused, frustrated, or ready to buy. Social Media Research Skills turns messy online chatter into clear business insights, revealing the specific words customers use to describe their problems. It also spots the competitor ads and viral posts actually driving results, so you know exactly what to copy and what to avoid. No more guessing what your audience wants; you get a clear, evidence-based picture of the market. Use this tool to stop guessing and start creating content that truly resonates with real people.

🆔 @hackernewsgithubprojects

📦 scenario-labs/blender-plugin

Scenario for Blender

Your AI agents can now build 3D scenes directly inside Blender. The Scenario for Blender plugin connects local coding agents to your open project through a Model Context Protocol server. This lets automated tools read your current scene state and generate meshes, materials, and video right into the viewport without manual file imports. It turns written descriptions into greybox layouts and applies generated assets instantly. It is essentially an API bridge between your creative workflow and automated code execution. If you want to script complex 3D production tasks using large language models, this is the direct path to making that happen.

🆔 @hackernewsgithubprojects

📦 fengzdadi/memory-bench-registry

memory-bench-registry

Track down exactly how every major A.I. memory test scores performance. The memory bench registry solves the problem of scattered, inconsistent testing data by collecting over two hundred benchmarks into one structured place. It works like a detective for research papers, tracing every single data point back to the original source to ensure you never guess a result. You get a clear list of every test that measures how well a machine remembers things over time. It filters out noise and only keeps verified, useful benchmarks. It stops you from wasting hours checking unreliable websites. This tool gives you a trustworthy map of the field.

🆔 @hackernewsgithubprojects

📦 atria-asi/atria-dawn

Atria Dawn: The AI Agent That Actually Finishes the Job

Atria Dawn is the AI agent that actually turns vague research questions into verifiable, finished results. Developed by the Shanghai Artificial Intelligence Laboratory, this massive model doesn't just chat; it works. It handles the entire workflow from analyzing a complex problem to writing the code, running the experiments, and even recovering when things go wrong. The real surprise here is its ability to act as a complete digital coworker. Instead of giving you a list of steps to do yourself, it executes the multi-step tasks for you, whether that means building a data visualization or drafting a technical report.

🆔 @hackernewsgithubprojects

📦 zehraakar424-lgtm/wallet-core-web

Xendra Wallet: Crypto in Your Browser

Xendra Wallet runs full crypto signing power directly in your browser using WebAssembly. This repository provides a lightweight interface for managing assets across multiple blockchains without requiring a heavy native app. You can import recovery phrases, swap tokens, and track balances right from your desktop or phone. The code is open for developers to inspect how trust-based cryptography works in a standard web environment. It is a fascinating look at bringing secure digital finance tools to the web platform. If you want to understand how modern wallets handle security without hardware, this project is worth a quick look.

🆔 @hackernewsgithubprojects

📦 antfu/furifont

Furifont: The Japanese Font That Reads Itself

Furifont is the Japanese font that displays pronunciation guides directly on the text itself. It solves the annoying problem of needing extra code or complex markup just to show how to read Japanese characters. By baking the pronunciation into the font file, the text stays completely plain and works instantly in any app that supports custom fonts. You just drop the font in, and the small reading text appears right above the main characters without any scripts. It is a clever way to make Japanese text accessible anywhere. If you work with Japanese web content, this tool saves you from heavy coding tasks.

🆔 @hackernewsgithubprojects

📦 tg-twilight/c001apk

C001APK: A Lightweight Coolapk Client

Browse a lighter, fully Chinese version of Coolapk without the official app. This unofficial client cuts out the bloat, letting you read feeds, check apps, and search topics with a simple, clean interface. It runs on Android and Windows, giving you control over themes and font sizes so the experience feels personal. The project actively fixes login issues and improves the browsing flow, making it a reliable choice for fans who want a faster, less cluttered way to follow their favorite tech content. If you are tired of heavy apps slowing you down, this is your lightweight alternative that respects your time and your device's battery.

🆔 @hackernewsgithubprojects

📦 minibandit/mkw-3ds

MKW-3DS: Mario Kart Wii on 3DS

Mario Kart Wii is running natively on the Nintendo 3DS. This open source project lets you build the game on your own computer directly from your own Wii disc, converting the original courses, characters, and sounds into a format that runs directly on the 3DS hardware. It is not an emulator, and the physics are calculated using a separate recreation of the original Wii math. The code is written specifically for the 3DS processor, meaning it is not just a scaled down version of the game. It supports Grand Prix, Time Trials, and even Balloon Battle with CPU opponents.

🆔 @hackernewsgithubprojects

📦 sakanaai/ai-scientist-v3

AI Scientist: Turn Tasks Into Research Papers

Turn a simple task description into actual scientific experiments and a finished research paper, driven entirely by coding agents. This repository orchestrates a full pipeline where AI proposes ideas, runs tests, analyzes results, and writes the manuscript. The coolest part is that it treats these coding tools as real scientists, handling everything from initial brainstorming to final peer review. It removes the massive bottleneck of manual data collection and paper writing, letting you focus on the actual discovery. If you want to see how far autonomous AI can push the research workflow, this is your blueprint. Stop manually formatting citations and start letting the agents do the heavy lifting.

🆔 @hackernewsgithubprojects

📦 xibushu/uniforge3d

UniForge3D: 3D Assets from Sparse Photos

UniForge3D is the 3D tool that builds complete digital objects from just a few scattered photos. It solves the tricky problem of turning unposed snapshots into a fully usable three-dimensional asset by combining reconstruction and generation in one shared space. This means you don’t need perfect camera angles or expensive sensors, just a handful of views to preserve fine details. It is a clever way to get high-quality 3D content without the usual heavy lifting, making it a genuinely useful resource for anyone exploring spatial computing.

🆔 @hackernewsgithubprojects

📦 google-research/harness2

Harness2: The AI That Fixes Itself

Watch your AI agent get better in real time. Google Research’s Harness2 doesn’t just run your model; it actively rewrites the system around it. Think of it as a coach that watches your AI work, spots exactly where it stumbles, and instantly adjusts its instructions, tools, and memory to fix the mistake. It then remembers that lesson for the next task. This creates a loop where your agent gets smarter with every job, even without changing the core brain. It is a fascinating way to make AI adaptable to new, unexpected challenges.

🆔 @hackernewsgithubprojects

📦 tencentarc/player-1

Player-1: Teaching AI to Play from Synthetic Experience

Player-1 teaches AI models to control video games by letting them practice on synthetic experience rather than watching humans play. The core idea is simple: instead of relying solely on human demonstrations, the system first fine-tunes a computer-use model on curated gameplay data, then sends it into a gym of sixteen synthetic environments where it interacts with verifiable rewards to build real experience. This approach helps the agent learn general game control—aiming, driving, navigating, and using interfaces—across diverse 2D and 3D scenarios.

🆔 @hackernewsgithubprojects

📦 seerray-lab/xiaomi-ocr-0

Xiaomi OCR Zero: Tiny Model, Big Reads

Xiaomi OCR Zero is the tiny language model that finally reads your messy papers with a giant brain. It is a small, efficient engine that turns images of receipts, handwritten notes, and dense math equations into clean, usable text without needing a cloud server. The real surprise is that this compact model matches the performance of much larger systems, handling everything from ancient calligraphy to modern spreadsheets with high accuracy. You get a local tool that understands what you see, not just what is written. It is a powerful proof that you do not need massive hardware to get professional-grade reading results on your own machine.

🆔 @hackernewsgithubprojects

📦 phuryn/work-humanizer

Work Humanizer: The AI Tone Remover

Work Humanizer is a specialized tool that rewrites professional text to sound like a person instead of a robot. The most surprising part is not just the rewrite, but the rigorous proof that it works. The creator used blind judges and hand-checked every single output to guarantee that the meaning stayed exactly the same while the robotic flavor vanished. This project solves a very real problem. When you write with AI assistance, your words can start to feel generic and cold. This tool strips away that artificial feel so your emails, posts, and memos sound like you.

🆔 @hackernewsgithubprojects