📦 microsoft/mage

Mage

Microsoft just released a four-billion parameter model family called mage that matches the quality of AI systems five to eight times its size. Instead of throwing raw computing power at image generation and editing, this project co-designs its image tokenizer and model backbone to focus detail exactly where the visual signal is. This means you can generate high-resolution images or perform complex, instruction-based image edits with massive speedups on ordinary hardware, and it even renders clean bilingual text. It is a fantastic, lightweight setup for developers who want top-tier results without renting a massive server cluster. Check it out to run fast local generation.

🆔 @hackernewsgithubprojects

📦 tanishq-dubey/macos-laguna-s2.1

macos-laguna-s2.1

macos-laguna-s2.1 is the local benchmark harness that finally makes it easy to find the absolute fastest and most accurate way to run the massive Laguna coding model on your Mac. If you are experimenting with local code assistants, you know that picking the right model size and format is usually a guessing game. This tool automatically downloads, runs, and evaluates different compressed versions of the model against a suite of real-world python coding tasks. It tracks memory use, loading speed, and actual token output directly on apple silicon, revealing exactly which setup gives you the best performance without sacrificing intelligence.

📰 https://news.ycombinator.com/item?id=49001323

🆔 @hackernewsgithubprojects

📦 lspr98/conformer-stt-s3

Conformer STT S3

Run voice recognition directly on a tiny microcontroller without ever connecting to the internet. The conformer-stt-s3 project brings a compressed, thirteen-million parameter English speech-to-text model straight to the ESP32-S3 chip. Instead of sending your private voice data to a massive cloud server, this system processes and transcribes your speech completely on the device, ensuring total privacy. By using clever math shortcuts, custom processor instructions, and split-core processing, it squeezes a heavy deep-learning model into just a few megabytes of memory. It is a massive win for building secure, low-power, and battery-friendly smart home gadgets that work anywhere.

🆔 @hackernewsgithubprojects

📦 atomicbot-ai/atomic-agent

atomic-agent

You can now run a highly capable AI agent completely offline on your own device without sacrificing advanced reasoning or tool usage. Built specifically for local models, atomic-agent is a local-first companion that operates entirely on your hardware to ensure absolute privacy. Unlike standard setups that struggle with complex actions offline, this project introduces custom tool-calling grammars that force local models to use system tools, search the web, and manage files reliably. It features a built-in terminal interface, long-term memory consolidation, and deep integration with the Model Context Protocol. It is the perfect playground for building private, highly capable assistants that run anywhere.

🆔 @hackernewsgithubprojects

📦 dingwu1021/silent-failures-multimodal-agentic-search

Silent Failures in Multimodal Agentic Search

Expose the hidden errors in visual search assistants before they lead your applications astray. When advanced AI agents search the web using both text and images, they often produce correct final answers despite completely ignoring the image, hallucinating sources, or contradicting the visual evidence. The silent-failures-multimodal-agentic-search project introduces a diagnostic framework to catch these invisible blunders. By running full agent search paths through a specialized evaluation judge and stress-testing them with blank images, this tool measures true correctness rather than surface accuracy. It is a fantastic way to understand how your visual search models actually behave under the hood so you can build more reliable agentic systems.

🆔 @hackernewsgithubprojects

📦 happynessi/prefix_grpo

Prefix GRPO

Train small-model AI agents more effectively by slicing and reusing successful teacher trajectories rather than relying on basic imitation. The prefix_grpo repository introduces a clever approach to reinforcement learning by splitting a teacher's step-by-step rollout into replayable starting points, restoring those exact environmental states, and training the student model on how to continue successfully from those mid-game moments. By optimizing both the historical context tokens and the ongoing actions, the system helps smaller models learn complex reasoning tasks in environments like text games and grid worlds. It provides the experimental code, modified trainers, and validated datasets to help developers teach smaller models to act like expert agents.

🆔 @hackernewsgithubprojects

📦 salmon1802/unirank

UniRank

UniRank completely changes how we test recommendation AI by ditching the lazy standard practice of only predicting a user's final action and replacing it with a rigorous step-by-step chronological simulation. In real life, recommendation models often get evaluated on messy, mismatched datasets. This open benchmark levels the playing field by testing fifteen major ranking models from tech giants like Google, Meta, and ByteDance across the exact same five massive industrial datasets. By standardizing the evaluation, it finally answers crucial questions about how model size, history length, and architecture structure affect real-world accuracy and hardware efficiency.

🆔 @hackernewsgithubprojects

📦 gnipbao/story-to-handdrawn-video

story-to-handdrawn-video

This clever tool converts plain Chinese story text or a simple sequence of uploaded images into a hand-drawn vertical diary-comic animation. It automatically splits your text into dynamic narrative beats, generates matching artwork, and runs a beautiful left-to-right visual reveal that transitions from written text to black-and-white sketch, and finally into a full-color illustration. Built on Remotion, it delivers a silent, beautifully framed video with optional paper-curl page turns that is perfectly structured for you to record a voiceover on top. It is the ultimate automation shortcut for turning raw story scripts into engaging, stylized social media content.

🆔 @hackernewsgithubprojects

📦 deer-flow/llm-space

LLM Space

llm-space is the desktop playground that finally takes the guesswork out of building and debugging AI agents. If you have ever tried coding an agent, you know how incredibly frustrating it is when the model calls and tools run in a black box and silently fail. This app runs locally on your machine, giving you a visual timeline to trace every single model call and tool execution as it happens. You can literally replay failed runs step by step, tweak your prompts, and watch the agent improve. It is the ultimate local workbench for turning raw agent ideas into working code.

🆔 @hackernewsgithubprojects

📦 mdx-tom/gpt-5.6-instruct

gpt-5.6-instruct

Your AI assistant is probably refusing to help with complex tasks like reverse engineering or security testing because it thinks they are too risky. The gpt-5.6-instruct repository solves this by framing these technical challenges as safe, local sandbox tasks. It is essentially a specialized toolkit that lets you deploy customized instructions to bypass unnecessary AI refusals while keeping your workflows clean and organized. It features an interactive command-line tool to quickly preview, install, or rollback different versions without messing up your configuration. It is a brilliant way to make your local model actually do what you ask.

🆔 @hackernewsgithubprojects

📦 asz798838958/freeagentidentity

freeagentidentity

freeagentidentity is the local desktop panel that completely automates creating and managing free ChatGPT accounts. Built with Python and Electron, this handy application lets you handle bulk registrations right from your computer without dealing with tedious manual verification loops. It handles your proxies, email routing, and captcha solvers behind the scenes while running multi-threaded registration tasks automatically. You get a clean web dashboard to monitor your account lists, track execution logs, and export your newly generated credentials instantly. It is a brilliant way to manage your testing credentials in one secure dashboard.

🆔 @hackernewsgithubprojects

📦 persiyanov/herdr-reviewr

herdr-reviewr

Comment directly on an AI agent's code modifications without ever leaving your terminal. This clever companion integrates beside your terminal chat, letting you review syntax-highlighted diffs, select lines, and write comments. With a single keystroke, you can send all your feedback straight back to the agent as a structured list. It also includes a read-only pull request viewer, file search, and customizable color themes. It is a brilliant way to guide your AI coder through tricky tasks. Check out herdr-reviewr to make collaborative terminal-based coding a breeze.

🆔 @hackernewsgithubprojects

📦 ericrollei/comfy_hunyuanimage3

comfy_hunyuanimage3

comfy_hunyuanimage3 is the ComfyUI integration that finally lets you run Tencent's massive eighty-billion parameter image generator right on your local hardware. Normally, a model of this scale is completely out of reach for consumer setups, but this project changes everything by introducing smart quantization options like four-bit and eight-bit precision. It manages to shoehorn this giant AI onto single-GPU setups and even twenty-four gigabyte cards using clever memory budgeting and offloading. Beyond just fitting the model on your machine, it introduces memory-efficient expert routing that lets you render massive high-resolution images without crashing your system. It is the ultimate local playground for high-end open-source image generation.

🆔 @hackernewsgithubprojects

📦 coppsary/motionly

Motionly

Build and customize stunning motion graphics using a smart AI assistant that outputs real, editable project files. Motionly gives you a sleek visual canvas and timeline editor where you can drag, scale, and adjust keyframes, but underneath it all is a clean human-readable code format. Instead of spitting out a closed-off video file, the optional AI assistant drafts the actual animation markup so you can visually tweak every single detail, transition, and audio track yourself. It is the perfect bridge between prompt-based creation and precise timeline control. Grab your graphics, load the editor, and make something cool.

🆔 @hackernewsgithubprojects

📦 helldez/bigmoeonedge

BigMoeOnEdge

You can now run a massive sixty gigabyte artificial intelligence model on a standard phone with only twelve gigabytes of memory. A clever project called bigmoeonedge makes this possible by storing the bulk of the model directly on your phone's flash storage and only loading the specific experts needed for each word as they are requested. By bypassing typical memory limitations and utilizing parallel read lanes, it delivers usable generation speeds completely offline on a plain processor without requiring any special graphics hardware. It is a brilliant way to run giant models on everyday mobile hardware. You can check out their open-source code and download the demo app to try...

📰 https://news.ycombinator.com/item?id=48969917

🆔 @hackernewsgithubprojects

📦 kbrdn1/gwm-cli

gwm-cli

You can now restore a deleted Git worktree without digging through your reflog. This capability comes from gwm-cli, a single-binary Git worktree manager written in Rust that works directly through native operations instead of slow shell commands. It completely automates the tedious setup process by automatically copying configuration files, running your build setup, and linking the active GitHub issue the second you create a branch. When you are done, a built-in terminal interface lets you manage all your active workspaces, run commands across them simultaneously, and even open your terminal shell or Git tools directly inside the application. It is a fantastic way to keep your development environment tidy without...

🆔 @hackernewsgithubprojects

📦 zsniko/e-specformer

e-specformer

An ultra-lightweight artificial intelligence model can now run complex radio frequency analysis directly on a Raspberry Pi in less than a single millisecond. Known as e-specformer, this network replaces the heavy mathematical operations standard in typical transformers with a linear attention mechanism that completely cuts out high-overhead functions. Built for real-time spectrum monitoring, it processes raw signal samples directly at the extreme edge. This design achieves over ninety-four percent accuracy on hardware Trojan detection while running on tiny devices like field-programmable gate arrays. This project makes sophisticated radio signal analysis practical for incredibly small, low-power hardware.

🆔 @hackernewsgithubprojects

📦 rockychen0205/omnireasoner

OmniReasoner

Most AI models completely choke on long videos because processing hours of high-definition footage at once is just too expensive. OmniReasoner solves this by teaching the AI to act like a smart video editor. Instead of squinting at a giant, blurry preview, it does a quick, cheap scan first, realizes it needs more detail, and literally calls a built-in zoom-in tool to grab a high-quality snippet of the exact seconds it needs. It is like giving an AI a magnifying glass and a remote control, letting it choose when and where to look closer before giving you a highly accurate answer.

🆔 @hackernewsgithubprojects

📦 recursive-knowledge/ksi

ksi

Improve how AI agents solve complex coding and reasoning tasks by letting them learn from their own mistakes. Instead of running a single agent and hoping for the best, ksi deploys a whole population of temporary agents in sandboxed containers to tackle your tasks. After working independently, they join a structured forum to compare notes on what worked and what failed. The system distills their collective wisdom into reusable guidance that seeds the next generation of agents, saving this knowledge in a shared store so it survives across runs. It is an incredibly clever way to make AI self-improve without training a new model.

🆔 @hackernewsgithubprojects

📦 grignarder/uvfacefusion

UVFaceFusion

Researchers have built a system that can reconstruct a complete, matching 3D digital model of a human face from ordinary photos in under three seconds. The project, called uvfacefusion, takes multiple images of a person captured from different angles in everyday settings and merges them into a single consistent 3D mesh. Instead of relying on slow, complex scanning equipment, it uses a smart neural fusion technique in 2D texture space to combine the details quickly. It makes high-quality 3D face modeling incredibly fast and accessible, allowing anyone to generate an accurate digital head with just a few clicks.

🆔 @hackernewsgithubprojects