This media is not supported in your browser
VIEW IN TELEGRAM
📦 nemocake/claude-obsidian-assistant
Automate Your Obsidian Notes with Claude Code
This Obsidian vault template integrates directly with Claude Code to automate project documentation and note management. By utilizing a structured Johnny Decimal folder system, the setup allows Claude to read your codebase, track development logs, and generate visual architecture diagrams as native canvas files. You can sync active projects, update project metadata, or refresh key documentation files using simple commands. This framework provides a flexible foundation meant for customization, enabling you to tailor workflows, themes, and automated task tracking to your specific needs. Start by cloning the repository to turn your notes into an active development workspace.
🆔 @hackernewsgithubprojects
Automate Your Obsidian Notes with Claude Code
This Obsidian vault template integrates directly with Claude Code to automate project documentation and note management. By utilizing a structured Johnny Decimal folder system, the setup allows Claude to read your codebase, track development logs, and generate visual architecture diagrams as native canvas files. You can sync active projects, update project metadata, or refresh key documentation files using simple commands. This framework provides a flexible foundation meant for customization, enabling you to tailor workflows, themes, and automated task tracking to your specific needs. Start by cloning the repository to turn your notes into an active development workspace.
🆔 @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
📦 meetopenbot/openbot
OpenBot: A Local-First OS for AI Agents
OpenBot is a local-first platform designed to orchestrate and run multiple AI agents on your own hardware. By providing a dedicated event-driven server, it solves the challenge of agent fragmentation by allowing different specialized agents and plugins to communicate seamlessly through a shared bus. The system uses a modular architecture where you can define custom agents using markdown files and extend capabilities with plugins like bash or memory tools. It handles everything locally, keeping your threads and configuration in your own file system. It is a powerful way to build and manage a private, event-driven agent ecosystem today.
📰 https://news.ycombinator.com/item?id=48645700
🆔 @hackernewsgithubprojects
OpenBot: A Local-First OS for AI Agents
OpenBot is a local-first platform designed to orchestrate and run multiple AI agents on your own hardware. By providing a dedicated event-driven server, it solves the challenge of agent fragmentation by allowing different specialized agents and plugins to communicate seamlessly through a shared bus. The system uses a modular architecture where you can define custom agents using markdown files and extend capabilities with plugins like bash or memory tools. It handles everything locally, keeping your threads and configuration in your own file system. It is a powerful way to build and manage a private, event-driven agent ecosystem today.
📰 https://news.ycombinator.com/item?id=48645700
🆔 @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
📦 clisystems/utfs
Store Named Data on Embedded Devices Without a Full File System
Manage structured data on raw flash or EEPROM by treating it like a series of named files using this lightweight storage layer. Designed for resource-constrained microcontrollers, the project provides a TAR-inspired way to store data blobs back-to-back with tiny headers, avoiding the complexity and overhead of full file systems. It uses only two source files, requires zero dynamic memory allocation, and lets you define your own read and write functions to match your hardware. By registering your structs and using a simple load-and-save lifecycle, you can easily persist application settings across power cycles while keeping your firmware footprint minimal.
📰 https://news.ycombinator.com/item?id=48571007
🆔 @hackernewsgithubprojects
Store Named Data on Embedded Devices Without a Full File System
Manage structured data on raw flash or EEPROM by treating it like a series of named files using this lightweight storage layer. Designed for resource-constrained microcontrollers, the project provides a TAR-inspired way to store data blobs back-to-back with tiny headers, avoiding the complexity and overhead of full file systems. It uses only two source files, requires zero dynamic memory allocation, and lets you define your own read and write functions to match your hardware. By registering your structs and using a simple load-and-save lifecycle, you can easily persist application settings across power cycles while keeping your firmware footprint minimal.
📰 https://news.ycombinator.com/item?id=48571007
🆔 @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
📦 latent-spaces/brag
Turn Your Projects Into Viral Launch Videos With One Command
Brag is a specialized Claude Code skill designed to instantly transform any software project into a professional, twenty-second launch video. By running a single command within your project directory, you trigger an automated process that generates custom music, motion graphics, and shareable copy tailored to your work. This tool removes the tedious burden of manual video editing, allowing builders to produce high-quality demo content that is ready for social media immediately. It streamlines the entire marketing workflow for developers, ensuring your latest creation gets the attention it deserves with minimal effort. Simply install the skill and let your code speak.
🆔 @hackernewsgithubprojects
Turn Your Projects Into Viral Launch Videos With One Command
Brag is a specialized Claude Code skill designed to instantly transform any software project into a professional, twenty-second launch video. By running a single command within your project directory, you trigger an automated process that generates custom music, motion graphics, and shareable copy tailored to your work. This tool removes the tedious burden of manual video editing, allowing builders to produce high-quality demo content that is ready for social media immediately. It streamlines the entire marketing workflow for developers, ensuring your latest creation gets the attention it deserves with minimal effort. Simply install the skill and let your code speak.
🆔 @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
📦 johell1ns/browser-search
Automate Web Browsing for AI Agents
This system provides a free, self-hosted toolkit that enables AI agents to autonomously search and navigate the web by orchestrating three specialized open-source tools. It uses SearXNG for efficient multi-source searching, Camofox to handle standard browsing through a REST API, and CloakBrowser to bypass aggressive anti-bot protections like Cloudflare. By automating this escalation process, the agent chooses the right tool for the task without human intervention. This setup runs on minimal hardware like a Raspberry Pi, allowing agents to perform deep research, cross-verify information, and extract structured data from protected websites while remaining completely free and unlimited.
🆔 @hackernewsgithubprojects
Automate Web Browsing for AI Agents
This system provides a free, self-hosted toolkit that enables AI agents to autonomously search and navigate the web by orchestrating three specialized open-source tools. It uses SearXNG for efficient multi-source searching, Camofox to handle standard browsing through a REST API, and CloakBrowser to bypass aggressive anti-bot protections like Cloudflare. By automating this escalation process, the agent chooses the right tool for the task without human intervention. This setup runs on minimal hardware like a Raspberry Pi, allowing agents to perform deep research, cross-verify information, and extract structured data from protected websites while remaining completely free and unlimited.
🆔 @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
📦 mayukh4/linux-android
Turn Old Androids Into Linux Desktops
An arm64 Android phone with at least 3 gigabytes of RAM can be transformed into a functional Linux desktop or a smart home server using this repository. By leveraging Termux and Termux-X11, these scripts automate the installation of graphical environments like XFCE4 or MATE, alongside tools like Firefox and Wine, without requiring root access. It enables GPU-accelerated performance on devices with Snapdragon chips while providing a way to manage home automation via Home Assistant Core. This project effectively repurposes old hardware into capable computing devices, giving you a powerful mobile workstation or a dedicated server right from your pocket.
🆔 @hackernewsgithubprojects
Turn Old Androids Into Linux Desktops
An arm64 Android phone with at least 3 gigabytes of RAM can be transformed into a functional Linux desktop or a smart home server using this repository. By leveraging Termux and Termux-X11, these scripts automate the installation of graphical environments like XFCE4 or MATE, alongside tools like Firefox and Wine, without requiring root access. It enables GPU-accelerated performance on devices with Snapdragon chips while providing a way to manage home automation via Home Assistant Core. This project effectively repurposes old hardware into capable computing devices, giving you a powerful mobile workstation or a dedicated server right from your pocket.
🆔 @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
📦 raketenkater/ggrun
Stop manually tuning GGUF models
Launch GGUF models on complex multi-GPU setups without manually configuring tensor-split or KV-cache flags. This tool automatically detects your hardware, including RAM and PCIe topology, to compute the optimal model placement and backend choice for llama.cpp or ik_llama.cpp. It features a smart terminal interface for downloading hardware-matched models, benchmarking performance with its AI Tune engine, and recovering from crashes automatically. By balancing expert placement across heterogeneous GPUs, it often achieves significantly faster generation speeds compared to default settings. Use this to simplify your local LLM deployment and squeeze every bit of performance from your hardware.
📰 https://news.ycombinator.com/item?id=48647221
🆔 @hackernewsgithubprojects
Stop manually tuning GGUF models
Launch GGUF models on complex multi-GPU setups without manually configuring tensor-split or KV-cache flags. This tool automatically detects your hardware, including RAM and PCIe topology, to compute the optimal model placement and backend choice for llama.cpp or ik_llama.cpp. It features a smart terminal interface for downloading hardware-matched models, benchmarking performance with its AI Tune engine, and recovering from crashes automatically. By balancing expert placement across heterogeneous GPUs, it often achieves significantly faster generation speeds compared to default settings. Use this to simplify your local LLM deployment and squeeze every bit of performance from your hardware.
📰 https://news.ycombinator.com/item?id=48647221
🆔 @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
📦 blazejkustra/react-native-morph-view
Smooth Gooey Image Morphing in React Native
Native mobile applications can now achieve fluid shape-morphing transitions using a specialized view component that blends images with a gooey metaball effect. Instead of a standard crossfade, this library employs alpha-threshold shaders on both iOS and Android to fuse silhouettes together as they transform. Developers control the transition using a single toggle property, while tunable parameters like blur radius allow for adjusting the molten appearance of the animation. By leveraging the latest React Native architecture, this tool provides a high-performance solution for creating liquid, interactive UI elements that feel polished and inherently connected across different screen states.
🆔 @hackernewsgithubprojects
Smooth Gooey Image Morphing in React Native
Native mobile applications can now achieve fluid shape-morphing transitions using a specialized view component that blends images with a gooey metaball effect. Instead of a standard crossfade, this library employs alpha-threshold shaders on both iOS and Android to fuse silhouettes together as they transform. Developers control the transition using a single toggle property, while tunable parameters like blur radius allow for adjusting the molten appearance of the animation. By leveraging the latest React Native architecture, this tool provides a high-performance solution for creating liquid, interactive UI elements that feel polished and inherently connected across different screen states.
🆔 @hackernewsgithubprojects
Media is too big
VIEW IN TELEGRAM
📦 dtquocbao/ai-market
Build Smarter Trading Strategies with this Multi-Agent AI System
Analyze financial markets using a sophisticated multi-agent AI system that combines real-time data from Yahoo Finance with advanced machine learning techniques. This platform utilizes ten specialized agents to deliver technical analysis, reinforcement learning strategy optimization, and risk management through an Angular dashboard. Key capabilities include LLM-powered event analysis, sector-specific ticker discovery, and dynamic A/B testing to refine trading performance. By integrating PostgreSQL for reliable data persistence and a robust agent routing system, it transforms complex market data into actionable insights and forecasts. Dive into the repository to explore how automated agent orchestration can enhance your market intelligence.
🆔 @hackernewsgithubprojects
Build Smarter Trading Strategies with this Multi-Agent AI System
Analyze financial markets using a sophisticated multi-agent AI system that combines real-time data from Yahoo Finance with advanced machine learning techniques. This platform utilizes ten specialized agents to deliver technical analysis, reinforcement learning strategy optimization, and risk management through an Angular dashboard. Key capabilities include LLM-powered event analysis, sector-specific ticker discovery, and dynamic A/B testing to refine trading performance. By integrating PostgreSQL for reliable data persistence and a robust agent routing system, it transforms complex market data into actionable insights and forecasts. Dive into the repository to explore how automated agent orchestration can enhance your market intelligence.
🆔 @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
📦 shadowhackrs/gmail-account-creator
Automated Gmail Creation with Advanced Anti-Detection
This Python automation tool streamlines bulk Gmail account creation by combining sophisticated anti-detection systems with intelligent bypass strategies. It mimics human behavior through randomized typing delays, session warming, and unique browser fingerprints for every account. The software integrates seamlessly with external services for phone verification and uses proxy rotation to ensure reliable operations. With a modern, console-based interface, users can track real-time progress and success rates, making it an efficient solution for managing multiple accounts. This project ultimately provides a structured and automated way to navigate complex registration flows while maintaining a low detection profile.
🆔 @hackernewsgithubprojects
Automated Gmail Creation with Advanced Anti-Detection
This Python automation tool streamlines bulk Gmail account creation by combining sophisticated anti-detection systems with intelligent bypass strategies. It mimics human behavior through randomized typing delays, session warming, and unique browser fingerprints for every account. The software integrates seamlessly with external services for phone verification and uses proxy rotation to ensure reliable operations. With a modern, console-based interface, users can track real-time progress and success rates, making it an efficient solution for managing multiple accounts. This project ultimately provides a structured and automated way to navigate complex registration flows while maintaining a low detection profile.
🆔 @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
📦 team-attention/tide
Tide: The Open-Source Workbench for AI Agents
Tide is a macOS application designed as an open-source alternative for developers who want to move beyond simple chat interfaces when working with AI agents. It functions as a unified workbench that consolidates your agent sessions, code files, terminal commands, and browser previews into one cohesive workspace. By keeping all project context and agent interactions on a single surface, the tool prevents the fragmented workflow typical of juggling multiple browser tabs and windows. Whether you are using Codex, Claude, Gemini, or other local agents, Tide helps you manage your software tasks more effectively and stay in the flow.
🆔 @hackernewsgithubprojects
Tide: The Open-Source Workbench for AI Agents
Tide is a macOS application designed as an open-source alternative for developers who want to move beyond simple chat interfaces when working with AI agents. It functions as a unified workbench that consolidates your agent sessions, code files, terminal commands, and browser previews into one cohesive workspace. By keeping all project context and agent interactions on a single surface, the tool prevents the fragmented workflow typical of juggling multiple browser tabs and windows. Whether you are using Codex, Claude, Gemini, or other local agents, Tide helps you manage your software tasks more effectively and stay in the flow.
🆔 @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
📦 jaywebtech/autoshorts
Automate Vertical Video Clips with AI
Transform long-form video or audio recordings into high-impact, vertical short-form clips using this local-first desktop application. The tool automates your entire production pipeline by importing media, transcribing content with Deepgram, and using AI models like DeepSeek or Claude to rank the most viral potential moments. It leverages FFmpeg to handle native portrait auto-cropping, ensuring your final assets are ready for social media. By managing your projects and data locally, this software simplifies the workflow from raw footage to finished clips. Take control of your content creation process and turn your long recordings into engaging short-form highlights effortlessly.
🆔 @hackernewsgithubprojects
Automate Vertical Video Clips with AI
Transform long-form video or audio recordings into high-impact, vertical short-form clips using this local-first desktop application. The tool automates your entire production pipeline by importing media, transcribing content with Deepgram, and using AI models like DeepSeek or Claude to rank the most viral potential moments. It leverages FFmpeg to handle native portrait auto-cropping, ensuring your final assets are ready for social media. By managing your projects and data locally, this software simplifies the workflow from raw footage to finished clips. Take control of your content creation process and turn your long recordings into engaging short-form highlights effortlessly.
🆔 @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
📦 agentsmill/age-of-agents
Turn Your AI Coding Sessions Into a Pixel-Art Kingdom
Visualize your AI coding workflow by transforming every session into a peaceful, living pixel-art realm. This tool monitors your coding activity and maps it to a virtual environment where each active session appears as a settler, tasks act as workshops, and processed tokens fill your storehouse. By running alongside your terminal, it offers a relaxing, glanceable view of your development progress in either a fantasy or sci-fi setting. It bridges your command line and a game-like dashboard to provide a unique, calm perspective on your work. Keep your coding sessions organized while enjoying a creative, gamified display of your productivity.
🆔 @hackernewsgithubprojects
Turn Your AI Coding Sessions Into a Pixel-Art Kingdom
Visualize your AI coding workflow by transforming every session into a peaceful, living pixel-art realm. This tool monitors your coding activity and maps it to a virtual environment where each active session appears as a settler, tasks act as workshops, and processed tokens fill your storehouse. By running alongside your terminal, it offers a relaxing, glanceable view of your development progress in either a fantasy or sci-fi setting. It bridges your command line and a game-like dashboard to provide a unique, calm perspective on your work. Keep your coding sessions organized while enjoying a creative, gamified display of your productivity.
🆔 @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
📦 koki7o/claude-code-for-beginners
Master AI Development with Claude Code in 15 Modules
This repository provides 15 structured modules designed to help you master AI-powered development using Claude Code. As a project-based course, it guides you through everything from basic terminal setup and file manipulation to complex tasks like debugging, Git operations, and production deployment. By engaging in these self-paced lessons, you will learn how to effectively communicate coding requirements to AI agents, manage background processes, and integrate tools to build and ship real-world software applications. Whether you are a total beginner or an experienced developer, this course offers the practical skills needed to transform your coding workflow today.
🆔 @hackernewsgithubprojects
Master AI Development with Claude Code in 15 Modules
This repository provides 15 structured modules designed to help you master AI-powered development using Claude Code. As a project-based course, it guides you through everything from basic terminal setup and file manipulation to complex tasks like debugging, Git operations, and production deployment. By engaging in these self-paced lessons, you will learn how to effectively communicate coding requirements to AI agents, manage background processes, and integrate tools to build and ship real-world software applications. Whether you are a total beginner or an experienced developer, this course offers the practical skills needed to transform your coding workflow today.
🆔 @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
📦 kostard/forgecad-public-kit
Master Code-First CAD with ForgeCAD
ForgeCAD provides a comprehensive public toolkit for building code-first parametric models using TypeScript and JavaScript. This companion repository serves as a centralized hub for developers to access ready-to-run examples, installable agent skills, and detailed documentation for the ForgeCAD platform. By defining mechanical parts and assemblies as clean, parameter-driven code, it solves the challenge of rigid and opaque CAD workflows, allowing users to automate design iterations or integrate AI agents directly into their modeling process. Whether you are debugging geometry or building complex assemblies, this toolkit offers the necessary assets to refine your CAD logic and streamline engineering.
🆔 @hackernewsgithubprojects
Master Code-First CAD with ForgeCAD
ForgeCAD provides a comprehensive public toolkit for building code-first parametric models using TypeScript and JavaScript. This companion repository serves as a centralized hub for developers to access ready-to-run examples, installable agent skills, and detailed documentation for the ForgeCAD platform. By defining mechanical parts and assemblies as clean, parameter-driven code, it solves the challenge of rigid and opaque CAD workflows, allowing users to automate design iterations or integrate AI agents directly into their modeling process. Whether you are debugging geometry or building complex assemblies, this toolkit offers the necessary assets to refine your CAD logic and streamline engineering.
🆔 @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
📦 virgiliojr94/book-to-skill
Turn Any Technical Book into a Powerful Agent Skill
Convert your technical books, documentation, and research into specialized agent skills that you can directly query while you work. Instead of wasting time digging through PDFs or dealing with hallucinated answers, this tool processes your files into structured mental models, glossaries, and cheatsheets. These skills integrate seamlessly with tools like Claude Code and GitHub Copilot, allowing you to load specific chapters on demand and reference complex frameworks without constantly exhausting your token budget. Stop struggling to recall buried information and turn your local knowledge into an interactive, reliable resource that helps you code and study more effectively.
📰 https://news.ycombinator.com/item?id=48649087
🆔 @hackernewsgithubprojects
Turn Any Technical Book into a Powerful Agent Skill
Convert your technical books, documentation, and research into specialized agent skills that you can directly query while you work. Instead of wasting time digging through PDFs or dealing with hallucinated answers, this tool processes your files into structured mental models, glossaries, and cheatsheets. These skills integrate seamlessly with tools like Claude Code and GitHub Copilot, allowing you to load specific chapters on demand and reference complex frameworks without constantly exhausting your token budget. Stop struggling to recall buried information and turn your local knowledge into an interactive, reliable resource that helps you code and study more effectively.
📰 https://news.ycombinator.com/item?id=48649087
🆔 @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
📦 every-app/open-seo
Build Your Own SEO Platform
Developers can now self-host a full-featured SEO suite that integrates directly with DataForSEO and Google Search Console to provide professional analytics. This project functions as an open-source alternative to commercial tools like Semrush and Ahrefs, offering capabilities such as comprehensive site auditing, backlink analysis, rank tracking, and keyword research. Built with a modern tech stack, it provides modular server functions and a React-based interface for managing SEO projects and tracking performance trends. By allowing users to host their own infrastructure, it offers a transparent, flexible way to monitor search visibility and technical health. Explore the codebase to customize your own optimization workflow.
🆔 @hackernewsgithubprojects
Build Your Own SEO Platform
Developers can now self-host a full-featured SEO suite that integrates directly with DataForSEO and Google Search Console to provide professional analytics. This project functions as an open-source alternative to commercial tools like Semrush and Ahrefs, offering capabilities such as comprehensive site auditing, backlink analysis, rank tracking, and keyword research. Built with a modern tech stack, it provides modular server functions and a React-based interface for managing SEO projects and tracking performance trends. By allowing users to host their own infrastructure, it offers a transparent, flexible way to monitor search visibility and technical health. Explore the codebase to customize your own optimization workflow.
🆔 @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
📦 tao-11-chen/forehoi
ForeHOI: 3D Object Reconstruction from Hand-Object Videos
Reconstruct 3D object geometry and estimate poses directly from casual, daily hand-object interaction videos using ForeHOI. This feed-forward framework solves the difficult problem of severe hand-object occlusion by simultaneously predicting 2D mask inpainting and 3D shape completion. It eliminates the need for complex pre-processing and delivers results in under one minute, offering a significant speedup over traditional optimization-based techniques. By integrating 2D and 3D information, the model effectively processes monocular footage to produce detailed 3D meshes and accurate pose overlays. Explore this efficient, high-fidelity approach to capturing 3D objects from your everyday video recordings today.
🆔 @hackernewsgithubprojects
ForeHOI: 3D Object Reconstruction from Hand-Object Videos
Reconstruct 3D object geometry and estimate poses directly from casual, daily hand-object interaction videos using ForeHOI. This feed-forward framework solves the difficult problem of severe hand-object occlusion by simultaneously predicting 2D mask inpainting and 3D shape completion. It eliminates the need for complex pre-processing and delivers results in under one minute, offering a significant speedup over traditional optimization-based techniques. By integrating 2D and 3D information, the model effectively processes monocular footage to produce detailed 3D meshes and accurate pose overlays. Explore this efficient, high-fidelity approach to capturing 3D objects from your everyday video recordings today.
🆔 @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
📦 douglasmonsky/codex-usage-tracker
Track Your Codex Token Usage Locally
Analyze your local Codex token and usage credit consumption with a private, dashboard-driven tool that indexes your session logs into SQLite. This project helps developers identify which specific threads, models, or subagents are driving costs by providing a local dashboard, CLI reports, and companion skills for conversational insights. It keeps sensitive prompts, assistant messages, and secrets off your machine while highlighting patterns like low cache reuse and context bloat. Gain visibility into your usage patterns without ever uploading logs to the cloud, giving you complete control over your development data and a clearer picture of your ongoing efficiency.
🆔 @hackernewsgithubprojects
Track Your Codex Token Usage Locally
Analyze your local Codex token and usage credit consumption with a private, dashboard-driven tool that indexes your session logs into SQLite. This project helps developers identify which specific threads, models, or subagents are driving costs by providing a local dashboard, CLI reports, and companion skills for conversational insights. It keeps sensitive prompts, assistant messages, and secrets off your machine while highlighting patterns like low cache reuse and context bloat. Gain visibility into your usage patterns without ever uploading logs to the cloud, giving you complete control over your development data and a clearer picture of your ongoing efficiency.
🆔 @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
📦 renning22/glm-5.2-4090
Run GLM-5.2 on Consumer RTX 4090 GPUs
Twenty-four RTX 4090 GPUs can now run the 753-billion parameter GLM-5.2 model using native FP8 precision. This project provides a portable port of the DeepSeek-Sparse-Attention kernel stack, which previously required hardware restricted to enterprise-grade architectures like Hopper. By using a drop-in script, these kernels are successfully adapted for Ada Lovelace hardware, enabling interactive speeds of roughly 10 tokens per second on commodity cards. This implementation includes kernel verification, tuned GEMM configurations for better performance, and MoE fixes. It effectively allows users to run massive state-of-the-art models on hardware previously considered incompatible, proving the viability of high-end AI inference at home.
🆔 @hackernewsgithubprojects
Run GLM-5.2 on Consumer RTX 4090 GPUs
Twenty-four RTX 4090 GPUs can now run the 753-billion parameter GLM-5.2 model using native FP8 precision. This project provides a portable port of the DeepSeek-Sparse-Attention kernel stack, which previously required hardware restricted to enterprise-grade architectures like Hopper. By using a drop-in script, these kernels are successfully adapted for Ada Lovelace hardware, enabling interactive speeds of roughly 10 tokens per second on commodity cards. This implementation includes kernel verification, tuned GEMM configurations for better performance, and MoE fixes. It effectively allows users to run massive state-of-the-art models on hardware previously considered incompatible, proving the viability of high-end AI inference at home.
🆔 @hackernewsgithubprojects