📦 hermannbjorgvin/clawdmeter

Clawdmeter: Your Desk Dashboard for AI Usage

Watch your Claude Code usage climb right on your desk. Clawdmeter is a tiny ESP32 dashboard that connects over Bluetooth to your computer and displays your session and weekly API limits. The coolest part is the pixel art mascot, Clawd, which gets busier and more frantic as you burn through your quota, giving you a funny visual cue when you are getting heavy. It even doubles as a keyboard, sending space and tab shortcuts directly to your app so you can control voice mode without touching the mouse.

🆔 @hackernewsgithubprojects

📦 neelm0906/mference

Run Huge AI Models on 8GB Macs

Mference lets you run massive artificial intelligence models on small Apple computers by keeping only the active parts in memory while streaming the rest from your hard drive. It uses Swift and Metal to pull data directly from your solid-state drive, allowing huge systems with billions of parameters to work on just eight gigabytes of random access memory. You get a simple app to chat with large language models or a server you can connect to locally without uploading your data to the cloud. It is a clever way to keep your private conversations private while enjoying powerful computing on everyday hardware.

🆔 @hackernewsgithubprojects

📦 makazhanalpamys/soup

Train an 8B Model on a 4GB Laptop

You can actually fine-tune a massive eight billion parameter AI model on a laptop with just four gigabytes of graphics memory. That sounds impossible until you see how this tool called Soup works. Instead of trying to load the entire model into your video memory all at once, which usually crashes your machine, Soup uses a clever technique called layer streaming. It feeds the AI model to the GPU one piece at a time while keeping the heavy frozen parts sitting in your regular system RAM. This means you can train powerful models right on your own hardware without needing expensive cloud servers or specialized data center cards.

📰 https://news.ycombinator.com/item?id=49166984

🆔 @hackernewsgithubprojects

📦 yhfgyyf/vllm-deepseek-v4-sm89

Run DeepSeek V4 Flash on RTX 4090 with vLLM

This is the vllm-deepseek-v4-sm89 project, the hardware bridge that finally lets you run the massive DeepSeek V4 Flash model on standard RTX 4090 graphics cards using vLLM. Most people think this powerful AI requires expensive enterprise server chips, but this repository patches the underlying code to translate complex neural network operations into instructions your consumer GPU can actually understand. It effectively unlocks the full speed of this advanced language model on hardware you can buy today. By bridging the gap between high-end model architecture and accessible hardware, it proves you do not need a data center to experiment with state-of-the-art artificial intelligence.

🆔 @hackernewsgithubprojects

📦 neuroaihub/brainpilotbench

BrainPilotBench

BrainPilotBench turns scientific research into a fair test. It gives artificial intelligence agents four real neuroscience jobs, like analyzing brain scans or decoding neural signals, then grades them on the actual code and figures they produce instead of just how they behave. This removes guesswork by using strict, pre-set rules to score every submission, ensuring that only systems that truly understand the science get top marks. It is the first benchmark where you can trust the results because the scoring is handled by maintainers who protect the test data from leaks.

🆔 @hackernewsgithubprojects

📦 leonickson1/swiftlet

Swiftlet: Run Giant AI on Your iPhone

Swiftlet is the open-source runtime that lets you run massive, billion-parameter artificial intelligence models on ordinary Apple devices, including standard iPhones. Most large language models require gigabytes of RAM to stay active, but Swiftlet takes a clever shortcut by keeping only a small, essential part of the model in memory and streaming the rest directly from your storage chip just in time. This means you can chat with huge language models on your phone without needing a powerful computer or an internet connection. It is a brilliant example of optimizing code to run beautifully on everyday hardware.

📰 https://news.ycombinator.com/item?id=49158333

🆔 @hackernewsgithubprojects

📦 atlascloudai/awesome-seedance-2.5-prompts-skills

Awesome Seedance 2.5 Prompts Skills

This repository packs over a hundred carefully chosen video prompts for Seedance 2.5, each paired with a real video preview so you actually see what works. It goes further by offering an installable agent skill that helps you plan, optimize, and build storyboards before the video even starts, turning vague ideas into clear shots. You get reliable workflows for everything from simple single clips to complex sequences with consistent characters and products. The project explains how to stitch scenes together, handle transitions, and fix common drift without guessing. It is a practical guide that saves hours of trial and error.

🆔 @hackernewsgithubprojects

📦 mingti-org/phyai

PhyAI: Fast AI for Robots

PhyAI lets robots think and react in real time by running heavy AI models directly on tiny edge devices like NVIDIA Jetson boards. It solves the frustrating lag that usually makes robotic controls feel clumsy or unresponsive. Instead of waiting for slow cloud servers, this framework squeezes out maximum speed using specialized coding tricks and smart memory management. It even supports fancy data compression techniques to keep things fast without losing accuracy. Whether you are testing on a single chip or scaling to huge server clusters, it handles both smoothly. This is a game changer for building robots that feel truly responsive.

🆔 @hackernewsgithubprojects

📦 osilly/vision-deepresearch

Vision-DeepResearch: AI That Actually Looks Around

Teach an AI to actually look around before answering by letting it search the web dozens of times instead of just guessing. This project trains multimodal models to treat images as starting points for long, iterative investigations. It combines visual understanding with hundreds of search engine queries to solve tricky questions that simple chatbots miss. The researchers also provide a benchmark to test how well these models handle this extended reasoning process. It is fascinating to see how adding search turns transforms a dumb image viewer into a curious investigator.

🆔 @hackernewsgithubprojects

📦 prism-shadow/gdpevo

GDP Evo: Testing AI Self-Evolution on Real Tasks

GDP Evo lets artificial intelligence agents learn and improve themselves by tackling actual business problems instead of just solving math puzzles. Think of it as a practice ground where AI gets handed real-world memos about things like warehouse shipments or monthly accounting closings. It watches how these digital assistants figure out the right steps to handle the paperwork and then measures whether they actually got better at the job over time. This is a big deal because most tests are too abstract to tell us if AI can truly handle the messy, complicated tasks we deal with every day.

🆔 @hackernewsgithubprojects

📦 jd-opensource/joyai-video-edit

Edit Live Video with JoyAI Video Edit

Watch JoyAI Video Edit transform a live video stream in real time using just a text instruction. Instead of waiting for an entire clip to finish or processing it in slow batches, this system edits frames the moment they arrive. It combines a large language model to understand your request with a diffusion engine that paints new pixels on the fly. The result feels like magic, letting you swap backgrounds, change clothing styles, or remove objects as you watch the screen. It achieves thirty frames per second, turning video editing from a tedious, offline chore into an instant, interactive experience that feels surprisingly natural.

🆔 @hackernewsgithubprojects

📦 xiaobin-rong/unipase

UniPASE Speech Enhancement

UniPASE transforms noisy audio into crystal-clear speech without adding weird robotic artifacts, a rare feat in generative models. This repository offers the official implementation of a state-of-the-art system designed to clean up voice recordings while keeping them sounding natural and authentic. It works by stripping away background noise and intelligently reconstructing missing sound details rather than just deleting bad parts. You get ready-to-use checkpoints that handle everything from basic clarity to fixing dropped audio packets. The setup is straightforward, requiring only standard Python libraries like PyTorch and SciPy to run inference on your own files.

🆔 @hackernewsgithubprojects

📦 fzkuji/gui-agent-harness

GUI Agent Harness

Give it a simple task and watch it operate your entire desktop. This project turns any AI into a GUI automation agent. You type a request like 'open the settings menu,' and it autonomously takes screenshots, spots buttons, and clicks for you. The cool part is its visual memory. It learns what interfaces look like after one look and remembers them later, so it gets faster every time. It even zooms in repeatedly to find tiny buttons you’d normally miss. It works on Mac, Windows, and Linux, using local or remote machines. It’s basically a digital assistant that can actually use a computer instead of just talking about it.

🆔 @hackernewsgithubprojects

📦 zero-mem/zero-mem

zero-mem: Peer-Reviewed AI Memory

Zero-Mem holds the official source code for a new memory system designed for large language models. Right now, the repository is quiet, showing only a promise that the full implementation and technical details will arrive after a rigorous peer review process. It does not offer code to download yet, nor does it explain the specific architecture or tricks it uses. Instead, it serves as a placeholder for researchers and developers waiting for a scientifically validated solution to model memory challenges. You can watch this space for updates, as the team is preparing to share their validated methods once the review is complete.

🆔 @hackernewsgithubprojects

📦 kirodotdev/kirocrew

Kiro Crew: The workspace that never forgets

Kiro Crew is the persistent workspace that never forgets a step. It acts as a memory for your development projects, storing task history and preferences so you never lose context when you log off. Instead of starting from scratch every time you return, Kiro Crew picks up exactly where you left off, learning from past mistakes to get smarter with each attempt. You can let it handle routine jobs unattended or chat with it through your favorite messaging apps. It runs locally on your own machine, keeping your data private while working hard in the background.

📰 https://news.ycombinator.com/item?id=49172209

🆔 @hackernewsgithubprojects

📦 zqxwce/vphone-ws

vPhone Workstation

Run a virtual iPhone right on your Mac without touching the command line. This native macOS app gives you a simple window to manage your research machines. You can browse existing setups, create new ones with a step-by-step guide, and boot the device with a single click. The coolest part is that the actual iPhone screen pops up in its own separate window, so you get a real Apple-like experience right on your desktop. It handles all the heavy lifting in the background, letting you focus on testing or experimenting with different iOS versions easily. Just install it and start exploring virtual Apple devices with zero hassle.

📰 https://news.ycombinator.com/item?id=49179328

🆔 @hackernewsgithubprojects

📦 ziangwu-77/reba

ReBA: Smarter Load Balancing for AI

ReBA is the smart load balancing technique that finally keeps vision and language tasks in perfect sync for large AI models. Standard AI models often struggle when switching between images and text, causing uneven workloads that slow everything down. ReBA fixes this by treating images and text as separate entities, ensuring that each part of the system gets a fair share of the work. It looks at how different types of data travel through the network and adjusts the routing so that no single part gets overwhelmed. This makes complex vision-language models faster and more efficient without needing extra hardware.

🆔 @hackernewsgithubprojects

📦 zhemdi/cadena

Turn Any 3D Mesh Into Editable Code

CADENA reverses engineering by turning a static 3D mesh into editable code. Instead of guessing the whole shape at once, it builds the model step by step, adding one operation like a cut or hole and checking the result against the original image. This stepwise approach ensures the final design is fully parametric and clean, solving the hard problem of reconstructing precise mechanical parts from rough scans. It effectively bridges the gap between visual data and functional engineering files for creators.

🆔 @hackernewsgithubprojects

📦 xiaomi-research/spatio-lm

SpatioLM: Spatial Intelligence for Vision-Language Models

SpatioLM lets AI models truly understand physical space in images and videos. It takes existing vision-language models and adds a lightweight module that teaches them about depth and 3D structure without needing special hardware. By learning from a teacher model that understands geometry, SpatioLM helps the AI answer questions like which object is closer or how things are arranged in a room. It works on standard photos and video clips, offering both reasoning skills and precise depth perception. This makes complex spatial tasks much easier for your favorite AI assistants to handle.

🆔 @hackernewsgithubprojects

📦 syhya/mlsys26-flashinfer-contest

How AI Agents Optimized CUDA Kernels for MLSys 2026

This is the MLSys 2026 FlashInfer Contest package that shows how AI coding agents can autonomously optimize complex computer chip instructions. It documents two winning approaches for generating high-speed graphics processing unit kernels. One method involves a human guiding the AI with specific skills, while the other lets an agent system work completely alone, planning and executing code without any human interference. The results are impressive, landing in the top three across multiple competition categories. You can see the exact step-by-step logs of how the AI solved these challenging coding tasks, offering a fascinating glimpse into the future of automated software engineering.

🆔 @hackernewsgithubprojects

📦 klboke/kkrepo

kkRepo: Your Self-Hosted Nexus Replacement

kkRepo lets you ditch the annoying limits of Sonatype Nexus by hosting every type of software package in one place. It supports Maven, npm, Python, Docker, Rust, Go, Helm, and many more, acting as a direct, compatible replacement for your existing setup. The real magic is the one-click migration tool that moves your users, permissions, and packages over without breaking a sweat, so your CI pipelines keep working exactly as they did before. It runs fast with low memory and handles large-scale storage via S3 or local files. If you are tired of component caps or slow upgrades, this is the clean, open-source fix you have been looking for.

📰 https://news.ycombinator.com/item?id=49179927

🆔 @hackernewsgithubprojects