📦 fabiotosi92/zipdepth

ZipDepth

You can now run state-of-the-art 3D depth estimation on a standard mobile phone in real time. This is possible thanks to zipdepth, an incredibly compact model that packs the power of massive AI foundation models into a tiny, six-million-parameter package. By learning directly from giant vision networks, it calculates the depth of any scene—from night driving to close-up textures—without needing any fine-tuning. It runs fast on everyday hardware, shifting effortlessly from high-end graphics cards to portable devices. For developers, this means you can build instant, zero-shot 3D depth mapping directly into on-device applications.

🆔 @hackernewsgithubprojects

📦 meigen-ai/opsd-v

opsd-v

opsd-v is the post-training framework that finally keeps long-form AI videos from falling apart after just a few seconds. Usually, when you generate long videos in short chunks, the AI relies on its own messy, half-baked history to write the next frames. Naturally, small errors build up until the motion freezes or turns to absolute mush. This project solves that by using real, clean video history as a smart teacher during training, while the model practices generating the next step. It learns to correct its own drift, giving you rock-solid, extended motion without slowing down the actual generation process at all.

🆔 @hackernewsgithubprojects

📦 hku-mmlab/uniclawbench

UniClawBench

UniClawBench is the evaluation benchmark that finally tests how smart AI assistants really are when left alone with a computer. Instead of just asking these digital agents simple questions, it drops them into a secure virtual environment to see if they can actually solve multi-step problems. It challenges them to browse the web, edit files, use desktop applications, and figure things out across four hundred distinct bilingual tasks. A hidden evaluator watches their moves and grades their success, while a visual dashboard lets researchers inspect exactly where the systems got stuck. It is a major step toward building AI that can safely handle real-world desktop chores.

🆔 @hackernewsgithubprojects

📦 tommasocerruti/linear-attention-architectures

linear-attention-architectures

The linear-attention-architectures repository is a research toolkit that makes scaling alternative language models incredibly simple by integrating modern linear attention mechanisms directly into a powerful training framework. Instead of relying on standard, resource-heavy attention, this project implements clever alternatives like DeltaNet and smart cross-layer routing strategies to pass key information smoothly through the model layers. It is built as a specialized fork of Megatron-LM, meaning you can easily train and test these fast architectures at a massive scale using ready-to-go SLURM launch scripts. If you want to experiment with highly efficient architectures without building the scaling infrastructure from scratch, this is your perfect playground.

🆔 @hackernewsgithubprojects

📦 lionsr/tnlean

tnlean

Researchers have successfully formalized the fundamental theorem of matrix product states in Lean 4. This math library, called tnlean, mathematically proves how different tensor representations generate the exact same physical quantum states. By translating complex quantum physics concepts into computer-verified proofs, it rigorously shows that these identical states are mathematically locked together by a change of basis. This project builds a solid foundation of computer-checked quantum information theory, including quantum channels and representations, ensuring absolutely zero room for human error in the math. It is a brilliant example of how modern coding tools are making cutting-edge quantum physics completely airtight.

🆔 @hackernewsgithubprojects

📦 cvsp-lab/moe-gs-studio

MoE-GS Studio

Researchers have figured out how to combine different 3D rendering algorithms into a single system that handles movement and motion far better than before. The repository moe-gs-studio serves as a central hub for this new approach, which uses a mixture of specialized AI experts to divide and conquer the complex task of rendering dynamic 3D scenes. Instead of relying on one massive, slow model to calculate how objects bend and move, this project coordinates multiple smaller, specialized models that work together in real-time. It is an exciting step forward that makes creating high-fidelity, movable digital worlds much more practical and efficient.

🆔 @hackernewsgithubprojects

📦 ugness/self-conditioned-fmlm

self-conditioned-fmlm

Generating high-quality text using continuous flow models usually requires dozens of slow, sequential steps, but a clever new approach achieves state-of-the-art text generation in just one or two steps. The project self-conditioned-fmlm implements a training technique called fixed-point flow distillation. Traditionally, models use self-conditioning to clean up their own drafts, but nobody quite understood how to optimize this for ultra-fast generation. This repository proves that self-conditioning is actually solving a mathematical fixed-point iteration, compressing that entire complex process into a highly efficient model. If you are experimenting with fast language generation, you can run this distillation pipeline to build incredibly fast text generators.

🆔 @hackernewsgithubprojects

📦 zhuqinfeng1999/awesome-panoramic-scene-analysis

Awesome Panoramic Scene Analysis

Awesome-panoramic-scene-analysis is the curated reading list that finally maps the entire landscape of 360-degree computer vision. Standard AI models struggle with panoramic photos because bending a sphere onto a flat screen introduces massive visual distortion that breaks normal spatial logic. This repository solves that by organizing over one hundred foundational papers into a clear guide on how to build neural networks that natively understand spherical geometry. It covers everything from adjusting basic visual filters to adapting massive modern foundation models for full-view understanding. It is the perfect starting point for any developer looking to master true all-around visual perception.

🆔 @hackernewsgithubprojects

📦 moore-developers/moore-wechat-article-downloader

Moore WeChat Article Downloader

Moore-wechat-article-downloader is the local-first tool that converts fleeting WeChat public articles into your own permanent, searchable knowledge base. Instead of letting great content get buried in your chat feeds, this smart assistant saves entire articles, high-quality images, and even reader comments directly to your computer as clean Markdown files. It is incredibly useful for content creators and researchers who want to analyze writing structures, track reader feedback, and feed clean data into AI models. You simply use everyday language to download specific links, sync your favorite accounts, or archive your personal bookmarks. Everything is stored privately on your hard drive, ready for your next big project.

🆔 @hackernewsgithubprojects

📦 enola-labs/enola

Enola

Map your entire codebase into a precise architectural blueprint for your AI coding assistant to use. This local tool parses your source code to build a deterministic knowledge graph of modules, data routes, and dependencies across multiple languages and repositories. Instead of letting an AI agent waste expensive tokens guessing how your files connect through trial and error, it provides concrete structural facts directly to your agent. This lets your assistant instantly compute complex queries, trace call paths, and evaluate the exact blast radius of a change before writing a single line of code. Connect it to your favorite developer tool to make your AI assistant truly understand your system...

🆔 @hackernewsgithubprojects

📦 mizzenai/hpsv3

hpsv3

The visual evaluation tool that finally judges AI images the way real humans do, hpsv3 uses a massive dataset of over one million image comparisons to rank picture quality and how well a graphic matches its text description. Instead of relying on rigid, outdated mathematical formulas, this system is trained on everything from vintage AI attempts to the latest high-end generator outputs and real photography. It lets you feed in your images and instantly receive a score reflecting true human visual preference. It is a fantastic way to objectively test which image generator performs best for your specific prompts.

🆔 @hackernewsgithubprojects

📦 blue-b/slopguard

SlopGuard

Triage AI-generated pull requests and issues automatically to save your sanity. This tool analyzes incoming contributions, tags where they came from, and slaps a quarantine label on them if they look like low-effort machine output. It runs static checks and can even use an AI judge to spot prompt injections, hallucinated bugs, or copy-pasted assistant responses. The best part is that it is completely human-in-the-loop and never auto-closes anything, so you always make the final call with a simple comment command. Check it out to keep your repository clean and focus on real human contributors.

🆔 @hackernewsgithubprojects

📦 phuryn/pm-brain

PM Brain

PM Brain is the markdown-based second brain that finally keeps your product management context organized in one local folder. Instead of scattering your strategy, user insights, and roadmaps across various cloud apps, this project lets you manage everything using plain text files on your laptop. It works directly with your AI assistant to read your active files, write updates, and even run a weekly sweep to flag drifting hypotheses or unresolved stakeholder concerns. Every decision is backed by a clear evidence trail, keeping your thought process entirely transparent. It is a simple, local, and incredibly organized system that keeps your product strategy sharp without the bloat.

🆔 @hackernewsgithubprojects

📦 davidondrej/skills

David Ondrej's AI Agent Skills

You can now give your AI assistants a massive upgrade using the exact same workflows developed by one of the community's top creators. The skills repository is a collection of reusable, highly practical building blocks designed for coding, research, and workflow agents. Instead of starting from scratch, you can load these pre-packaged instructions to help your agents self-schedule recurring tasks, coordinate parallel coding subagents, and even automate browser sessions natively on macOS. It solves the problem of unreliable agent behavior by providing structured, real-world patterns for everything from deep web research to advanced agent orchestration, making your AI assistants vastly more independent and capable.

🆔 @hackernewsgithubprojects

📦 ath-maas/pixelle-video

Pixelle-Video

You can now create fully finished, narrated videos with matching images and background music just by typing in a single topic. Pixelle-video is an automated short video engine that handles the entire creative pipeline for you. Once you feed it a prompt, the engine drafts an original script, designs custom AI illustrations or video clips, synthesizes voiceovers with optional voice cloning, and compiles everything into a polished, ready-to-share video. It is incredibly useful for creators who want to quickly visualizes ideas without touching complex editing software. Simply type in your concept, click generate, and watch your script transform into a finished story.

🆔 @hackernewsgithubprojects

📦 mbvlabs/andurel

Andurel

Andurel lets you generate a fully functional database-backed web application with a single command, automatically creating the models, controllers, user views, and routes you need. This web framework brings the legendary speed of Ruby on Rails straight to Go, giving you a complete toolset for rapid development. It combines high-performance tools like Echo, Tailwind CSS, and PostgreSQL, letting you choose between server-rendered HTML or reactive single-page apps. A dedicated development server automatically reloads your code and assets as you type, and when you are ready for production, another single command bundles everything into a fast static binary.

📰 https://news.ycombinator.com/item?id=48871211

🆔 @hackernewsgithubprojects

📦 kunal12203/webify-mcp

webify-mcp

The adaptive web research assistant that brings high-quality deep research to your local AI coding agents for a fraction of the usual cost. Instead of dumping massive pages of fifteen thousand tokens into your AI's context window, webify-mcp builds a smart structural map of web pages and surgically extracts only the exact HTML subtrees relevant to your query. It slashes your token usage by ninety-nine percent, turning massive documentation deep-dives into tiny, hyper-focused snippets of just a few hundred tokens. It plugs directly into your favorite editors to give your coding assistant instant, budget-friendly web intelligence.

🆔 @hackernewsgithubprojects

📦 strukto-ai/mirage

Mirage

Mirage is the unified virtual filesystem that lets AI agents interact with all your cloud services like a local hard drive. Instead of juggling a dozen different APIs, SDKs, and custom database integrations, you can mount tools like Slack, S3, Google Drive, and Postgres side-by-side. This means an AI agent that only knows basic bash commands can search, copy, and pipe data across completely different platforms using simple commands like grep and cp. It makes managing diverse cloud data incredibly easy, transforming how developers build and run smart, autonomous agents.

🆔 @hackernewsgithubprojects

📦 1weiho/open-slide

Open Slide

Build beautiful presentation decks using natural language by letting your AI coding assistant write the actual React code for you. With open-slide, you no longer have to manually wrestle with slide dimensions or restrictive templates. Instead, you describe what you want, and your AI agent automatically structures your slides on a fixed layout canvas. You can click any element directly inside your browser to leave a comment, and the agent will instantly apply those design edits for you. It is the perfect bridge between AI speed and custom web code, giving you professional presenter tools and static HTML exports.

🆔 @hackernewsgithubprojects

📦 yeqing17-2026/omniagent

OmniAgent

Write code, execute system tasks, and run web research with an AI assistant that actually learns from its mistakes in real time. OmniAgent introduces a brilliant dual-layer reflective loop that intercepts system risks instantly and converts execution failures into learned insights on the fly. Instead of burning through your tokens with massive prompts, it dynamically loads documents in graduated stages as your conversation deepens. Even better, it packs an industry-first four-layer dynamic security scanner that routes high-risk terminal commands through an interactive approval sandbox so your system stays safe. Grab it on GitHub and let your AI agent evolve with every single chat.

🆔 @hackernewsgithubprojects