298 subscribers
14K videos
14.6K links
Download Telegram
This media is not supported in your browser
VIEW IN TELEGRAM
πŸ“¦ shoestring-dev/strixfold

Run DeepSeek on Two Cheap AMD Boxes

Run a massive AI model on two affordable AMD computers without needing expensive NVIDIA hardware. Strixfold takes two Strix Halo machines and links them with standard USB4 cables to work together as one powerful system. It runs DeepSeek-V4.1-Flash, delivering about seventy-two percent of the speed of high-end NVIDIA rigs while producing identical results. The coolest part is how it splits the heavy math between both boxes over simple network cables, proving you don't need a data center to get serious performance. If you have spare hardware sitting around, this is your ticket to building a serious local AI setup. Check out the repo to see how it’s done.

πŸ†” @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
πŸ“¦ jackson-cmd/claude-multi-desktop

Claude Multi Desktop: Chat Across Your Accounts

Stop juggling your personal and work Claude accounts on Mac. With Claude Multi Desktop, you can run both side by side, each with its own sign-in, while sharing one single coding setup. The coolest part? Your sessions can literally talk to each other. You can ask your work session to check a list your personal session is managing, and it just happens. It is one simple script, no heavy software, and no mess. If you use two accounts, this is the clean way to stop the chaos and let your bots collaborate.

πŸ“° https://news.ycombinator.com/item?id=50520258

πŸ†” @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
πŸ“¦ kmic-68/llama.cpp

Llama dot cpp

Run large language models on your own computer with Llama dot cpp. This open-source project lets anyone execute artificial intelligence locally without needing a cloud server or massive new hardware. It is particularly interesting because it has been specifically tuned to make older graphics cards run these models much faster than standard versions, pushing performance nearly twice as high while keeping the math incredibly accurate. By handling the heavy lifting right on your machine, it gives you full control over your data and processing speed. You get a powerful, private AI assistant that works reliably on equipment you already own.

πŸ“° https://news.ycombinator.com/item?id=50146283

πŸ†” @hackernewsgithubprojects
Media is too big
VIEW IN TELEGRAM
πŸ“¦ yang-mills-group/openai-ym-20261028

OpenAI Proves Yang-Mills with Lean

Stop guessing and start proving. This project takes the legendary Yang-Mills problem and solves it using the Lean theorem prover, turning massive mathematical concepts into computer-checked logic. It doesn’t just claim a result; it builds a rigorous bridge from lattice calculations to physical reality, ensuring every step is mathematically perfect. For the first time, we have a machine-verified proof of this cornerstone of quantum physics, eliminating human error from the process. If you’ve ever wondered how we can be absolutely certain about the math behind the universe, this is your answer. It’s not just code; it’s certainty you can run.

πŸ†” @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
πŸ“¦ agentic-learning-ai-lab/deleting-weights

The deleting-weights

Teach a large language model new skills without touching nearly all of its weights. The deleting-weights project shows that zeroing out a tiny fraction of a model's internal connections can make it just as capable as a full, heavy update. This approach, called Deletion Tuning, solves a massive headache for developers who need to add features without losing what the model already knows. It keeps the original knowledge intact while the model learns new tasks, all while using a file size that is thousands of times smaller than standard updates. It is a clever way to make AI models smarter and leaner at the same time.

πŸ†” @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
πŸ“¦ anionex/h3-image-mlx

H3 Image MLX: Local AI Art on Your Mac

Generate and edit stunning images directly on your Mac using the H3 Image MLX library. It takes a popular AI model and runs it entirely on your computer’s built-in graphics chip, so you don’t need a cloud subscription or a supercomputer. The coolest part is how it manages memory. Instead of keeping the whole massive model loaded in your random access memory, it reads the data in tiny, on-demand chunks and releases it instantly. This lets you create huge, high-resolution photos using only a few gigabytes of space, even while your browser and other apps are running.

πŸ†” @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
πŸ“¦ budgetany/budget3d

Budget3D: One Model, Any Quality

Budget3D is the image-to-3D tool that lets you dial down quality without losing your mind. It takes a single photo and turns it into a spinning 3D object, but here is the cool part: you can tell it to use less computer power for the boring parts, like empty space, while keeping the details sharp where they matter. One model handles every setting, so you are not switching between different files for high or low res. It is like a volume knob for your 3D scene, letting you decide exactly how much detail you need.

πŸ†” @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
πŸ“¦ nvlabs/dvsm

DVSM: Real-Time 3D From a Single Transformer

DVSM is the view synthesis model that renders entirely new 3D scenes using just a single Transformer network. It solves the problem of needing a separate, slow optimization step for every new environment by reconstructing a scene and painting its pixels in one continuous pass. This is particularly exciting because it matches the quality of traditional methods that take much longer to process. You feed it a handful of photos, and it immediately lets you explore that space from any angle in real-time. It is a fast, self-contained way to build immersive 3D environments from simple flat images. This tool makes complex scene generation accessible and immediate.

πŸ†” @hackernewsgithubprojects
Media is too big
VIEW IN TELEGRAM
πŸ“¦ vincalkr/almond

Running Frontier AI on a Gaming GPU

Almond runs massive artificial intelligence models directly on standard consumer graphics cards. It handles complex language engines on a single or dual AMD Radeon setup. The software squeezes these heavy computational tasks into your existing video memory. It bypasses the need for expensive enterprise server hardware. This open source project brings high-end local inference to everyday creators. You can test powerful neural networks without renting cloud time. It democratizes access to advanced language generation tools. The engine is optimized specifically for consumer silicon performance. This is how you run next-gen AI on your desktop. No data center required, just your graphics card. This is Almond.

πŸ†” @hackernewsgithubprojects
Media is too big
VIEW IN TELEGRAM
πŸ“¦ spearsheep/anima-brain

Anima Brain: AI that actually remembers you

Your AI companion forgets everything the moment you close the tab. Anima Brain fixes that by giving it a local, private memory that tracks your facts, feelings, and even how your relationship evolves over time. Instead of just storing text, it connects past conversations to your current mood, so your chat partner feels consistent and caring rather than robotic. It runs entirely on your computer using a simple database, meaning your conversations stay yours and don't get sent to a distant server. If you want your AI friend to actually know you, this is the tool that makes it happen without the cloud headache.

πŸ†” @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
πŸ“¦ cobanov/pixelart-gaussiansplat

Turn 3D Models Into Pixel Art

Render real 3D scenes as crisp, retro pixel art right in your browser using pixelart-gaussiansplat. It takes standard 3D models and converts them into flat, colorful splats with hard outlines and a limited palette, so you get that classic retro look without painting every pixel by hand. The coolest part is that you can drop in any rigged character model, and it keeps its animations while the software automatically builds a floating island underneath to match its movement. It handles the tricky lighting and color snapping for you, turning complex geometry into clean, stylized art that is easy to orbit and explore.

πŸ†” @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
πŸ“¦ replit/datacat

DataCat

Run your own private observability stack without the cloud bill. DataCat lets you host metrics, logs, and traces directly on your servers, storing everything in databases you already run like PostgreSQL. The coolest part is that it speaks the same language as Datadog agents, so you can point existing tools at it without rewriting your code. It is self-contained, meaning no external accounts or hosted services are required to get insights. If you want full control over your data without sacrificing visibility, this is your new home base.

πŸ†” @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
πŸ“¦ krmanik/kioku

Kioku: Anki Desktop on Your Phone

Kioku runs the actual Anki desktop application on your smartphone, not just a simplified mobile clone. While the official mobile apps are great for basic reviews, they lack access to custom add-ons and the full card browser. Kioku solves this by packaging the unmodified desktop software to run natively on iOS and Android, or even in a browser tab. This means every power feature you rely on your computer, like advanced editors and specific add-ons, works exactly as intended right on your touch screen. It also lets you self-host the server if you want complete control over your data.

πŸ“° https://news.ycombinator.com/item?id=50207761

πŸ†” @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
πŸ“¦ panuvi/inclusionmed

InclusionMed: The New Standard for Medical AI

InclusionMed is the open evaluation initiative that finally checks if medical AI actually cares about real patients. It organizes tests by healthcare roles and AI capabilities to reveal exactly where models succeed and where they fail. By comparing twenty-two models on five specific benchmarks, it creates a clear scorecard that highlights gaps in coverage. This matters because a single high score means nothing if the tests ignore the diverse needs of people worldwide. It ensures progress in medical AI is defined by actual healthcare outcomes, not just raw data processing. If you are tracking the future of healthcare technology, this is the baseline you need to watch.

πŸ†” @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
πŸ“¦ atunfrito/tuna-fonts

Tuna Fonts: The Windows Font Manager That Actually Respects Your Privacy

Tuna Fonts is the privacy-focused font manager that finally lets you organize your Windows text without the cloud nonsense. It gives you a clear view of every font on your computer, letting you install, preview, and toggle them on or off without breaking your system. The coolest part is that you can browse and install almost two thousand Google Fonts with a single click, all while keeping your data completely local. It also lets you compare up to four fonts side by side so you can pick the perfect style for your project. No accounts, no ads, and no data sent anywhere.

πŸ†” @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
πŸ“¦ video-rebirth/comfyui-hyperflow-flash

HyperFlow Flash: Instant Video Generation

HyperFlow Flash is the ComfyUI plugin that runs massive video AI models on standard laptop hardware. It plugs into ComfyUI to let you generate high-quality videos using heavily compressed model files that fit in just a few gigabytes of memory. The coolest part is how it uses specialized math tricks to keep the video quality high while drastically cutting down the space and power needed to run it. You can generate text to video or reference to video clips in a few seconds without renting expensive cloud servers.

πŸ†” @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
πŸ“¦ isaacus-dev/cookbooks

Spot Contract Conflicts With AI

Detect legal conflicts between documents to catch hidden risks before they become problems. The Isaacus Cookbooks provide ready-made guides showing exactly how to run this process using their specialized AI tools. You can upload two contracts, and the system breaks them down, compares every single clause, and flags where the rules disagree or where a necessary term is completely missing. It turns a tedious manual review into an automated check that highlights deviations and gaps instantly. If you need to make sure two legal agreements actually align without reading them line by line, this is the practical way to do it.

πŸ†” @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
πŸ“¦ igawa6/peanuts-native

Peanuts Native: Play Switch Games on Android

Turn your Android handheld into a native Switch console with Peanuts Native. This project takes a game file you already own and rewrites its code to run directly on your phone, skipping the slow emulation layer entirely. The coolest part is how it handles dual-screen devices like the AYN Thor. The game plays on the top display, while the bottom screen shows a live, touchable companion. It pulls real-time data like maps and inventory straight from the running game, giving you a smart, interactive dashboard. It is a fascinating look at how native code can bridge the gap between handheld hardware and console games.

πŸ†” @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
πŸ“¦ braedonsaunders/ether

Ether: The Media Server That Holds No Media

Your family can press play on any movie in the entire database without you downloading a single file. That is exactly what Ether does. It is a media server that holds no media. Instead of filling your hard drive with copies, it talks directly to your existing streaming apps like Plex and Infuse. When you pick a title, it instantly grabs the exact video chunks from your debrid or Usenet accounts on demand. You get the full experience of a massive library with zero storage to manage. It is the cleanest way to watch anything, anytime, without the usual download mess.

πŸ†” @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
πŸ“¦ avi0n/meshcoreone

MeshCore One

MeshCore One is the offline messaging app that lets your phone talk directly to other devices without any internet connection. It builds a private radio network where messages hop from device to device, creating a resilient communication link that works even when cell towers and Wi-Fi are completely down. This is particularly useful for hikers, emergency responders, or anyone who needs to stay in touch in remote areas or during natural disasters. The app handles the complex radio protocols automatically, so you just open it, find your friends, and send messages like any other chat app.

πŸ“° https://news.ycombinator.com/item?id=50552948

πŸ†” @hackernewsgithubprojects