📦 don-chad/ninfer-3090

NInfer 3090

NInfer 3090 is the local inference engine that finally runs massive twenty-seven billion parameter artificial intelligence models directly on a single consumer graphics card. By writing highly optimized native hardware code from scratch, this system bypasses heavy frameworks to squeeze every drop of performance out of a standard graphics processing unit. It introduces custom memory layouts and clever batching techniques to let you serve complex reasoning models on your own machine without relying on expensive cloud servers. If you want to host private, incredibly fast language models on standard hardware, this project shows just how far consumer hardware can be pushed.

🆔 @hackernewsgithubprojects

📦 darksidewalker/dasiwa-comfyui-workflows

Dasiwa ComfyUI Workflows

Dasiwa ComfyUI Workflows is the workflow library that keeps your AI video generation setup perfectly organized through changing models. If you use ComfyUI, you know how fast things break when new models drop. This collection acts as a living archive of working setups for heavy hitters like Wan 2.2 and Lightricks LTX. Instead of rebuilding your complex image and video pipelines from scratch every single time, you can grab these pre-configured snapshots. It is a massive timesaver that lets you skip the tedious troubleshooting and get straight to making cool art.

🆔 @hackernewsgithubprojects

📦 zfu006/ssg

SSG: Self-Guided Image Generation

SSG is the frozen diffusion model that teaches itself to make better images using its own mistakes. Most AI image generators need a heavy external helper to guide their creativity, but SG proves that a fixed model can actually look at its own half-finished samples and use them to self-correct. This repository provides the official code and pretrained weights for that Synthetic Self-Guidance technique on pixel-based diffusion models. You can clone the repo, set up the Python environment, and download ready-to-use checkpoints. It lets you generate high-quality pictures directly from the command line without needing complex training runs.

🆔 @hackernewsgithubprojects

📦 compactifai/full-chunked-kl-loss

Full Chunked KL Loss: Shrinking AI Memory Needs

The full chunked KL loss kernel that slashes video memory usage by processing large language models in bite-sized pieces. This project offers a practical benchmark showing how chunking the training process avoids storing massive intermediate data, which traditionally crashes your GPU when sequences get long. Instead of computing everything at once, it processes small segments forward and backward, keeping your memory footprint tiny even with huge context windows. It is a clever engineering trick that lets you train complex AI models on smaller hardware without sacrificing accuracy.

📰 https://news.ycombinator.com/item?id=49257181

🆔 @hackernewsgithubprojects

📦 longlongaaago/ultradiffedit

UltraDiffEdit: Edit Images Up to 8K Resolution

UltraDiffEdit lets you edit massive, ultra-high-resolution photos without needing to train a new model from scratch. It takes standard AI tools and scales them up to handle images up to eight thousand pixels wide. The secret is a smart multiscale process that breaks the huge picture into smaller pieces, edits them one by one, and then blends everything back together seamlessly. This approach avoids the blurry seams and weird artifacts that usually ruin high-resolution edits. You can swap objects, change colors, or add details using simple text prompts or shape guides.

🆔 @hackernewsgithubprojects

📦 zhaobo253-crypto/attrguide

AttrGuide Ultrasound Classifier

AttrGuide injects clinical attribute priors into ultrasound classifiers. Instead of relying solely on raw visual texture, it uses a parallel branch to match image features with medical concepts like shape and margin. This dual-branch approach improves accuracy and provides interpretable evidence for why a diagnosis was made. The project offers a lightweight, plug-and-play framework that guides deep learning models with human-readable semantic cues, making computer-aided diagnosis more robust and transparent for medical imaging tasks.

🆔 @hackernewsgithubprojects

📦 opengeos/anymap-ts

anymap-ts

Build interactive custom maps inside Python notebooks using your favorite frontend mapping tools. This tool lets you render geographic data using MapLibre, Leaflet, Cesium, or even three-dimensional point clouds with Potree, all written in TypeScript under the hood. It sets up a bidirectional bridge so your Python data and your visual map stay perfectly in sync as you zoom, pan, or click. You can draw shapes, manage complex layers, and instantly export the final creation to a standalone HTML file. It is the ultimate superpower for anyone who loves python data but wants gorgeous, responsive web maps.

🆔 @hackernewsgithubprojects

📦 divagr18/memlayer

Memlayer

Large language models can now remember everything you tell them across multiple conversations with just three lines of code. A new tool called memlayer gives artificial intelligence agents persistent long term memory without any complex setup. It works like a smart filter, automatically ignoring greetings and filler words while extracting important facts, dates, and relationships. It then saves this information using a hybrid setup of vector databases and structured knowledge graphs, meaning the model can pull up relevant context in under one hundred milliseconds.

🆔 @hackernewsgithubprojects

📦 shivateam/driveteach-vla

DriveTeach-VLA

DriveTeach-VLA teaches self-driving models to recognize traffic scenes and plan paths by combining visual understanding with spatial guidance. It uses driving-aware vision distillation to help the model focus on key road elements, while trajectory-guided prompts align its predictions with real driving paths. This approach replaces complex reasoning with direct spatial grounding, making the system more reliable and precise. By integrating these techniques, the project offers a clearer way to improve autonomous driving performance through better visual learning and trajectory alignment.

🆔 @hackernewsgithubprojects

📦 neul-labs/grite

Grite: The Issue Tracker That Lives in Your Repo

Grite is the local issue tracker that lives inside your git repository, solving the messy problem of AI agents losing track of their work by keeping tasks and notes permanently tied to your code. Instead of relying on slow external servers or cluttering your files with messy notes, this tool stores every change as a secure, unchangeable log directly within git, allowing multiple coding assistants and developers to coordinate smoothly without stepping on each other's toes. You can create tasks, claim work, and store learning notes while staying completely offline, knowing that everything syncs automatically when you push changes.

🆔 @hackernewsgithubprojects

📦 nickjiang2378/value-axis

Value Axis: The Inner Compass of AI Models

The value axis project reveals that large language models actually encode a hidden signal indicating whether they are on the right track. Researchers identified a specific direction in the model's internal activity that measures task confidence. By steering the model along this axis, you can causally influence its behavior. Push toward high value and the model persists with its current approach. Push toward low value and it starts backtracking. This works consistently across math problems, coding tasks, and preference learning. The repository provides the tools to build and use this axis, letting you probe how well the model believes it is succeeding.

🆔 @hackernewsgithubprojects

📦 kuaishougamemind/trace-bench

TRACE BENCH: How AI Roleplay Testing Finally Gets Honest

TRACE BENCH is the evaluation platform that finally forces roleplay AI to prove it can handle complex, multi-step tasks instead of just making polite small talk. Most testing tools only check if a character stays in voice, but this tool assigns a smart digital interviewer to actively interrogate the model. This interviewer carries a private checklist and uses conversation tools to deliberately probe for specific behaviors, ensuring every requirement is actually met before signing off. It turns vague personality traits into hard, traceable scores. If you want to see which AI characters can truly deliver on their promises, this is the transparent testing ground you have been waiting for.

🆔 @hackernewsgithubprojects

📦 neur0map/glazepkg

GlazePKG

Manage every package manager on your computer with a single command. GlazePKG acts like a unified controller for over forty different package managers, including brew, pip, cargo, npm, and apt, across Linux, macOS, and Windows. Instead of memorizing dozens of different commands and flags for different systems, this tool lets you search, install, update, and roll back applications using one simple syntax. It features a beautiful, searchable terminal dashboard to view all your installed software in one place, take configuration snapshots, and even back up your entire setup to migrate to another machine. Keep your system clean and organized.

🆔 @hackernewsgithubprojects

📦 nicepkg/ctxport

CtxPort

Moving your deep AI conversations between different chatbots usually means dealing with broken formatting, missing code blocks, and a mountain of manual copying. CtxPort fixes this instantly by letting you grab any thread from ChatGPT, Claude, Gemini, or DeepSeek as a clean, structured Markdown file with just one click. What makes it incredibly clever is the sidebar hover feature, allowing you to instantly copy a thread directly from your chat history list without waiting for the actual page to load.

🆔 @hackernewsgithubprojects

📦 duelion/homebox-companion

Homebox Companion

Homebox Companion is the smart inventory assistant that turns photos of your physical belongings into organized digital data. Instead of typing out descriptions and serial numbers manually, you simply take a photo of a room, shelf, or box. The app uses advanced vision artificial intelligence to automatically identify multiple items in a single image, extracting details like the manufacturer, model, and retail price, and then saves them directly to your self-hosted setup. You can even use the built-in conversational assistant to find, move, or update your inventory just by chatting. It makes keeping track of your physical gear completely painless.

🆔 @hackernewsgithubprojects

📦 hustvl/eovsam

EOVSAM: Segmentation 300x Faster

EOVSAM is the AI segmentation model that finally cuts processing time by three hundred times without losing accuracy. Most systems struggle by guessing every object one by one, which is painfully slow. EOVSAM changes the game by generating all masks in a single pass using a new aggregation method that handles open-vocabulary tasks efficiently. This means it can identify and outline any object described in text instantly, rather than crawling through a list. It combines top-tier precision with incredible speed, making real-time visual understanding practical for the first time. If you want fast, accurate image analysis that scales with your needs, this is the breakthrough you have been waiting for.

🆔 @hackernewsgithubprojects

📦 parskatt/romav2

RoMa v2

Match points of interest between two different images with incredible accuracy, even under difficult lighting and perspectives. By building on top of deep vision models, this project estimates dense, highly accurate correspondences and maps overlapping areas across image pairs in a clean, pythonic way. This is perfect for computer vision tasks like reconstructing 3D scenes or registering images, as it bypasses traditional sparse feature tracking in favor of a robust, dense matching approach. Best of all, you can load a pretrained model and run it on your own images with just a few lines of code.

🆔 @hackernewsgithubprojects

📦 aagentah/nw_wrld

nw_wrld

Trigger jaw-dropping live visuals using web technologies you already know. With nw_wrld, you can code custom visual modules in plain JavaScript using libraries like p5 and Three.js, then orchestrate them in real time using a slick native interface. Launching the app opens two screens: a control dashboard and a clean output projector. You can choreograph your graphics using a built-in sixteen-step sequencer or map your visual methods to external triggers like MIDI keyboards, OSC signals, and live audio capture. It is the ultimate playground for interactive performances, installations, or prototyping, right from a portable project folder.

🆔 @hackernewsgithubprojects

📦 fasouto/termaid

Termaid

You can now render complete Mermaid diagrams directly in your terminal without spinning up a browser or calling an external web service. A clever new Python tool called termaid translates complex diagram syntax into beautiful, readable text art right in your console. It supports eighteen different diagram types, from flowcharts and sequence diagrams to git graphs and mindmaps. It even intelligently handles pie charts by converting them into clean, horizontal bar charts that actually make sense in text. With zero dependencies, built-in color themes, and automatic terminal-width fitting, it is the perfect utility for documenting systems over secure shell connections or inside terminal-based applications.

🆔 @hackernewsgithubprojects

📦 rrrrrrzy/fm-geometry

Free Uncertainty Detection with Flow-Matching Geometry

Accelerate your AI safety checks by reading the denoising path of a flow-matching model without running a single extra calculation. This project introduces a free proxy that spots uncertainty just by watching how the system bends during its own cleanup process. When the AI is certain, its internal path is smooth and straight, but when it is guessing or confused, the trajectory starts to wiggle. By measuring that physical curvature using basic math, you get an instant alarm before a failure happens. It works on everything from simple toy models to real robot arms, catching mistakes early while using zero extra computing power.

🆔 @hackernewsgithubprojects

📦 zihangtian/arco

ARCO: LLMs That Grade Themselves

ARCO is a Python project that lets language models grade their own step-by-step thinking without any external judges. Instead of waiting for a final score, the model generates a simple three-point checklist for every single action it takes, then scores that action against the checklist right away. Because the grading model and the problem-solving model train together on the same data, the checker learns to spot useful steps while the solver learns to earn better points. This creates a self-correcting loop where the agent improves its reasoning by constantly evaluating and adjusting its own process. It is a fascinating look at how models can become their own teachers.

🆔 @hackernewsgithubprojects