📦 modeltc/minimax-h3-turbo

Minimax-H3 Turbo: Generate Video in 4 Steps

Minimax-H3 Turbo makes high-quality video generation incredibly fast by cutting the required processing steps down to just four. This project distills the original heavy model into lightweight versions that run significantly quicker on your hardware. You can use these checkpoints with standard coding libraries or visual workflow tools, allowing you to create videos from text prompts or existing images with synchronized audio. The repository provides ready-to-use setups for both technical users who prefer code and creators who like visual interfaces. It handles various video resolutions and keeps the original style intact while speeding up the creation process.

🆔 @hackernewsgithubprojects

📦 duckyshell/comfyui-minimaxh3-prompt-writer

ComfyUI MiniMax H3 Prompt Writer

ComfyUI MiniMax H3 Prompt Writer lets you generate complex video prompts by mixing plain language with photos and clips. You describe your idea, add reference images or video clips, and the tool uses a local AI model to craft a detailed prompt for the MiniMax H3 video generator. It handles all the technical syntax so you just focus on the creative brief. The setup is flexible, supporting local models via Ollama or direct GGUF files, or remote APIs if you prefer. You get an editable result that you can refine or copy directly into your workflow.

🆔 @hackernewsgithubprojects

📦 shmimel/bee-write-back

Bee Write Back: The Dedicated Journaling Device

Write your thoughts into a private journal without letting your whole past come rushing back to haunt you. Bee Write Back is a clever Raspberry Pi device built around the WriterDeck form factor, designed specifically for nightly journaling. You open the app, answer a random prompt or just freewrite, and then lock it away until the next day. The beauty is in the separation: unlike a digital notebook that shows you every old entry, this system keeps your reflections fresh and focused on the present moment.

🆔 @hackernewsgithubprojects

📦 tonyd2wild/deepseek-v4-flash-dgx-spark

DeepSeek-V4-Flash on dual NVIDIA DGX Spark

Setting up multi-node serving on dual NVIDIA DGX Spark boxes usually ends in tears, with broken RDMA handshakes, cold-start timeouts, and silent driver bottlenecks. The deepseek-v4-flash-dgx-spark repository delivers a battle-tested recipe to run the DeepSeek-V4-Flash mixture-of-experts model split across two Spark nodes using vLLM. It orchestrates a tensor-parallel configuration over a two-hundred gigabit Ethernet RoCE link, utilizing FP8 quantization and multi-token prediction speculative decoding. The project includes a production-ready Docker image with native kernels, launch scripts, and solutions for the dreaded GID index mismatch. Sync your firmware, spin up the worker, and enjoy blazingly fast, long-context local serving.

🆔 @hackernewsgithubprojects

📦 carasibana/comfyui-h3-facerefine

Fix Blurry Faces in MiniMax H3 Video

MiniMax H3 struggles to render faces clearly when characters appear small in the frame. Comfyui-h3-facerefine solves this by automatically detecting each face, cropping it so it fills the screen, and letting the video model regenerate the details before stitching the result back into the original clip. This approach works because the model gets a close-up view instead of a distant blur, producing sharper expressions without changing the overall shot. The process runs frame by frame, tracking movement smoothly to keep the video stable. It is a practical way to fix a common generation flaw, giving you high-quality results even when the subject is far away.

🆔 @hackernewsgithubprojects

📦 igniteui/igniteui-mcp-testbed

Test if AI Actually Uses Your Tools

See if your AI agent actually uses the tools you gave it by running the Ignite UI MCP Testbed. It spins up a fresh, isolated container, scaffolds a web app, and hands it to an AI coding agent with specific MCP servers and skills enabled. The real magic is the History tab, which tracks exactly which tools the AI called, which skills it loaded, and—crucially—what it ignored. You can instantly see if your agent is ignoring your custom APIs or just ignoring everything. It solves the mystery of AI reliability by proving whether your setup actually works, not just whether the code compiles.

🆔 @hackernewsgithubprojects

📦 matlowai/comfyui-mainodes

ComfyUI MAINodes Fixes Smearing in Fast Motion

ComfyUI MAINodes adds two clever nodes to MiniMax-H3: a Motion Lab that stops fast-motion smearing by temporarily slowing down only the chaotic parts of a clip, and a Contact-Sheet node that generates five simultaneous views of a subject from one reference image. The Motion Lab works by identifying where the model struggles to keep up, stretching out those specific frames so the AI gets extra time to draw clear poses, and then snapping them back to normal speed afterward. It uses a smart oracle to spot the messy spots without needing extra training, leaving the stable background untouched.

🆔 @hackernewsgithubprojects

📦 donkeyusecorp/donkey

Donkey

Donkey is the local-first video editor that finally keeps your media safe on your own machine. Unlike cloud-heavy editors that upload your footage, Donkey runs its heavy lifting right inside your Mac. You get a full timeline with multi-track editing, captions, and effects, all processed locally so your files never leave your disk. The real magic happens with its AI assistant. You simply describe the shot you want, like a 1920s comic chase scene, and the app generates the video, voice, and music for you.

🆔 @hackernewsgithubprojects

📦 mesh2motion/mesh2motion-desktop

Mesh2Motion Desktop

Mesh2Motion Desktop is the offline desktop app that brings web animation tools to your local machine without the internet. Right now, it is a work in progress using the Tauri framework, which keeps the app tiny by packing Rust code into small binaries instead of heavy web standards. The core idea is wrapping the existing Mesh2Motion web experience into a native window, letting you browse and download animations directly from your file system. It currently saves files to downloads and includes developer tools, so it is still in early testing.

🆔 @hackernewsgithubprojects

📦 nathanw1014/strix-halo-llamacpp

Strix Halo llama.cpp: Blazing Fast Local AI

Strix Halo llama.cpp delivers massive speedups for AMD’s new Strix Halo graphics chips by bundling specialized fixes directly into a ready-to-run package. Most developers struggle with slow performance when running complex AI models on this specific hardware, but this project solves that by optimizing how the chip processes long contexts and quantized data. Instead of fighting with complicated build tools, you get a self-contained folder that includes the necessary GPU drivers, so you can start running large language models immediately. The result is a dramatic jump in processing speed, making it feasible to run heavy AI models locally without needing expensive server-grade hardware.

🆔 @hackernewsgithubprojects

📦 vivekchand/clawmetry

See Your AI Agent Think

Watch your AI agent think. ClawMetry installs once and instantly builds a live dashboard that tracks fourteen different coding assistants. Instead of guessing why your bot failed, you get a real-time flow diagram showing exactly how messages move from your prompts through the model and back. It auto-detects your setup and breaks down every token cost and tool call so you can stop flying blind. Stop guessing why your AI agent failed. ClawMetry installs once and instantly builds a live dashboard that tracks fourteen different coding assistants.

🆔 @hackernewsgithubprojects

📦 achrefelouafi/vegetationgeneratorthreejs

Watch Ivy Grow on Any Shape with Vegetation Generator

Here is a fully interactive procedural ivy generator that turns any 3D shape into a living plant host. You simply drag your mouse across a sphere, box, or custom model to paint a path, and the code instantly grows a realistic vine skeleton that follows your stroke. The real magic is the physics: each leaf acts like a tiny hinge that flutters realistically in the wind without ever detaching from the branch. You can tweak the stem thickness, branch density, and wind strength on the fly, watching the changes happen instantly. It even lets you bloom flowers by hovering over the vines.

🆔 @hackernewsgithubprojects

📦 dfki-av/egoforce

EgoForce: Live 3D Hand Tracking for VR

EgoForce is the computer vision tool that finally nails realistic 3D hand tracking in virtual reality using just a single head-mounted camera. Most current methods struggle with depth and scale issues, often requiring expensive, specialized datasets to train properly. EgoForce solves this by using your forearm position as a visual guide to stabilize the hand reconstruction, even when your arm is partially out of frame. It works seamlessly across different camera types, including fisheye lenses found in popular AR glasses. This means you get precise, real-time hand meshes that feel natural in immersive applications without needing bulky sensors or complex setups.

🆔 @hackernewsgithubprojects

📦 cutwire-studios/drift

Drift: The Free Video Editor That Reads Your Mind

Drift lets you magically isolate people from their backgrounds with a single click, turning complex video masking into a simple tap instead of a hours-long editing marathon. This free desktop editor uses that clever AI cutout feature alongside plenty of other handy tools like auto captions and cool visual effects. You simply drop in clips, tweak the look, and export without paying a dime or seeing any watermarks. It is a genuinely fun way to polish your videos quickly, making professional results accessible to everyone who just wants to make cool content.

🆔 @hackernewsgithubprojects

📦 mlnlp-world/paper-bibchecker

Paper-BibChecker: Spot Fake Citations Before You Submit

Verify your bibliography instantly by scanning BibTeX files against trusted academic databases to catch hallucinated references before submission. This tool cross-checks titles, authors, and identifiers using sources like OpenAlex and official conference proceedings to flag errors or non-existent papers. It is useful because it saves researchers from accidental mistakes and ensures credibility in peer review without manually hunting down every citation. You can run it locally or use the web interface to get a clear report on which entries are valid and which need attention. Always double-check the flagged items yourself, as the tool is designed to assist rather than replace your final judgment on reference accuracy.

🆔 @hackernewsgithubprojects

📦 bevel-software/hexis

Hexis: Central AI Skills and Knowledge

Hexis is the centralized workspace that manages your company’s AI skills and knowledge. It solves the messy problem of scattered context by storing everything in a git repository you own. Instead of copying prompts between tools, employees connect their AI agents once to this workspace. Then, every agent only sees the skills and data they are allowed to access. Managers stay in control through a simple review system for changes. It keeps your data private and lets you switch AI providers without rebuilding your tools from scratch.

📰 https://news.ycombinator.com/item?id=49260145

🆔 @hackernewsgithubprojects

📦 emollick/annals-kingdom

A Living Medieval Kingdom in a Single File

Annals Kingdom is the interactive history simulator that turns a single HTML file into a breathing medieval world with its own economy, dynasties, and wars. Instead of playing a traditional game, you watch a procedurally generated realm unfold like a nature documentary. Rivers carve valleys, towns grow and rot over centuries, and noble families scheme, marry, and clash in civil wars. The real magic is the self-writing chronicle that records every event, from famines to dragon appearances, creating a unique historical record for each seed. You can share a link to share the exact same world, letting friends explore the same rivers and royal grudges.

🆔 @hackernewsgithubprojects

📦 griffinwork40/agent-afk

Agent Afk

Kick off a complex coding task in your terminal and actually walk away. Agent Afk lets you run AI agents in the background so you can focus on life instead of staring at a loading bar. The real magic is in the transparency: every single step the AI takes gets logged into a clear, readable trace. You can check that record before anything ships to production, which means you never have to blindly trust the bot. If the agent gets stuck or needs your input, you get pinged. It’s like having a reliable pair programmer who actually waits for you when things get tricky.

🆔 @hackernewsgithubprojects

📦 jerber/arc-code

How an AI Agent Built Its Own Game Engine

Arc-code is the general-purpose AI harness that proves a language model can invent its own problem-solving tools instead of relying on pre-built ones. It runs the ARC puzzle benchmark using a standard coding assistant that starts with zero knowledge of the game. Instead of using a fixed grid calculator, the AI writes Python parsers and simulators from scratch for each unique level it encounters. It tests these scripts, fixes bugs, and uses them to navigate the puzzle. After finishing a game, the AI discards all that custom code and starts over with a blank slate for the next challenge.

🆔 @hackernewsgithubprojects

📦 nvidia/nemo-speech.cpp

NeMo-Speech.cpp: Fast C++ Speech AI

Run powerful voice AI on your own laptop with NeMo-Speech.cpp. This lightweight C++ runtime lets you transcribe, translate, and even synthesize speech using efficient models that actually work on standard hardware. It strips away the heavy dependencies usually required for machine learning, replacing them with a fast, portable engine built on the popular GGML framework. You get a simple command-line tool and a local web server that handles real-time audio processing without needing a cloud connection. It is genuinely cool because it turns your computer into a private speech engine that respects your privacy and runs surprisingly fast.

🆔 @hackernewsgithubprojects