Coding & Development

MiniMax-M2 is an open-source, compact, fast, and cost-effective Mixture-of-Experts (MoE) model designed for advanced coding and agentic workflows. With 230 billion total parameters and only 10 billion active parameters, it offers high performance in tasks like multi-file edits, coding-run-fix loops, and test-validated repairs, while maintaining powerful general intelligence. The model is engineered for end-to-end developer workflows and excels in agent performance, planning and executing complex, long-horizon toolchains across shell, browser, retrieval, and code runners. Its efficient design leads to lower latency, lower cost, and higher throughput, making it ideal for interactive agents and batched sampling. MiniMax-M2 is available via API and its weights are open-source for local deployment.

Colliding Scopes is a free, open-source web-based tool that transforms user-uploaded photos into dynamic kaleidoscope animations. Operating directly in the browser, it allows for real-time adjustments to animation speed, the number of kaleidoscope tiles, and canvas size. Users can easily export their creations as MP4 video files or save screenshots. The tool is designed for various creative applications, including generating Spotify canvas art, stylized video project animations, and marketing assets. It emphasizes client-side processing, ensuring user privacy as no images are stored or saved. Developed by Alan, it builds upon Luke Hannam's kaleidoscope algorithm, focusing on an intuitive front-end user interface and export functionalities.

NASLib is a modular and flexible framework designed to facilitate Neural Architecture Search (NAS) research by providing a common codebase to the community. It offers high-level abstractions for designing and reusing search spaces, along with interfaces to various benchmarks and evaluation pipelines. This enables researchers to implement and extend state-of-the-art NAS methods with minimal code. The library's modular nature allows for easy innovation on individual components, such as defining new search spaces while reusing existing optimizers, or proposing new optimizers with current search spaces. Developed by the AutoML Freiburg group, NASLib is continuously updated with new search spaces, optimizers, and benchmarks.

OpenAI Baselines offers a collection of high-quality, open-source implementations of various reinforcement learning algorithms. This project is designed to facilitate research by providing reliable baselines for comparison and further development. It supports Python 3.5+ and integrates with TensorFlow versions 1.4 to 1.14 (with a separate branch for TensorFlow 2.0). Researchers can use Baselines to train models for tasks like controlling MuJoCo humanoids or playing Atari games, with options for saving, loading, and visualizing trained models. The project emphasizes reproducibility and provides tools for logging and visualizing learning curves.

Top Contributors To Follow is a web-based tool designed to identify and showcase the most impactful users on Hugging Face. It provides a ranked table of model creators based on the cumulative likes their models have received within a selected month. Users can easily pick a specific month to see who the top contributors were during that period. Each entry in the table includes the user's name, their total likes for the chosen month, and quick links to their Hugging Face profile, making it simple to discover and follow leading figures in the AI community. This tool is particularly useful for those looking to identify influential creators, explore popular models, or stay updated on key contributors within the Hugging Face ecosystem.

StreamPETR is an official implementation of a research paper accepted by ICCV 2023, focusing on exploring object-centric temporal modeling for efficient multi-view 3D object detection. This open-source tool provides a robust framework for researchers and developers working in the field of computer vision and autonomous driving. Key features include support for StreamPETR, PETR, and Focal-PETR codebases, flash attention, deformable attention (RepDETR3D), and checkpoints. It also offers functionalities like sliding window training, efficient training in streaming video, TensorRT inference, and 3D object tracking. The repository provides detailed documentation for environment setup, data preparation, and training/inference procedures, along with model zoo results on NuScenes validation and test sets.

SegLossOdyssey is an open-source repository offering a comprehensive collection of loss functions specifically designed for medical image segmentation. This tool is invaluable for researchers and practitioners aiming to enhance the accuracy and robustness of their segmentation models, particularly in tasks involving highly imbalanced data. The collection includes implementations in PyTorch and Keras, covering a wide array of loss functions from various research papers and challenges. It highlights the effectiveness of compound loss functions for challenging segmentation tasks and provides a valuable resource for exploring and applying state-of-the-art loss functions in medical imaging.

DRL-Pytorch offers a comprehensive, open-source PyTorch implementation of numerous Deep Reinforcement Learning (DRL) algorithms. It provides a unified framework for popular methods such as Q-learning, Duel DDQN, Prioritized Experience Replay (PER), C51, Noisy DQN, PPO, DDPG, TD3, SAC, and ASL. Developers can easily train agents from scratch by navigating to the desired algorithm's folder and running the `main.py` script. The repository is designed for robustness and clarity, making it an excellent resource for researchers and practitioners looking to implement, customize, or compare different DRL approaches. It also includes recommended resources for DRL, such as simulation environments, books, online courses, and important research papers.

Lite YouTube Embed is an open-source custom element designed to significantly improve the performance of embedded YouTube videos on websites. It renders videos approximately 224 times faster than a traditional YouTube iframe, focusing on visual performance and quicker loading times. The tool uses `youtube-nocookie.com` for enhanced user privacy and supports progressive enhancement for deferred loading with JavaScript. Developers can customize poster images, access the YouTube Iframe Player API, add video titles, and apply custom player parameters to control video behavior and appearance. It is available as an npm package and can be easily integrated by including its CSS and JavaScript files.

Daily Dictation English provides a comprehensive platform for English language learners to enhance their listening, writing, and speaking skills through interactive dictation exercises. The website features thousands of audio recordings and videos across various topics, including short stories, daily conversations, and specialized content for TOEIC, IELTS, and TOEFL exams. Users engage in a four-step process: listening to audio, typing what they hear, checking and correcting errors, and reading aloud for pronunciation practice. The platform caters to all levels from basic to advanced, offering a 100% free experience to improve English proficiency quickly and effectively.

compromise is an open-source JavaScript library designed to simplify natural language processing tasks. It provides core functionalities for analyzing text, breaking it down into tokens, and identifying parts of speech. The library's primary goal is to make NLP more accessible and straightforward for developers to integrate into their applications, focusing on modest NLP requirements rather than complex, large-scale models.

Learnable-triangulation-pytorch is an official PyTorch implementation of the paper "Learnable Triangulation of Human Pose" (ICCV 2019, oral). This open-source project focuses on 3D human pose estimation from multiple cameras, offering two novel methods: Algebraic and Volumetric learnable triangulation. These methods significantly outperform previous state-of-the-art techniques, with the Volumetric model achieving a 2.4 times reduction in error. The repository provides code for training and evaluation, supports both single and multi-GPU setups, and includes pretrained models and configurations for the Human3.6M dataset. It is designed for researchers and engineers working on advanced computer vision tasks, particularly in human pose estimation.

Deep_Metric is an open-source project offering PyTorch implementations for various deep metric learning methods. It is specifically designed to facilitate research and development in image retrieval and other information retrieval applications. The repository features implementations of prominent loss functions such as Contrastive Loss, Semi-Hard Mining Strategy, Lifted Structure Loss, Binomial BinDeviance Loss, NCA Loss, and Multi-Similarity Loss. Notably, it includes the code for XBM (Cross-Batch Memory), which was nominated as a best paper at CVPR 2020, demonstrating significant improvements in recall on large-scale datasets. The project also provides processed datasets like CUB and Cars-196 to aid in easy reproduction of experimental results, making it a valuable resource for researchers and practitioners in the field.

Model Comparator Space Builder is an AI tool designed for comparing various AI models. It provides a platform for researchers and data scientists to effectively evaluate the performance of different models and benchmark their results against each other. This tool is instrumental in the model selection process, helping users make informed decisions based on comparative analysis. It supports research and development efforts by offering a structured environment for model assessment, which is crucial for advancing AI applications. The tool aims to streamline the process of understanding model strengths and weaknesses, contributing to more robust and efficient AI solutions.

YOLO26 vs RF-DETR is a Hugging Face Space designed for comparing the performance of two prominent object detection and segmentation models: YOLO26 and RF-DETR. Users can upload an image and then choose between detection or segmentation tasks. The tool provides options to adjust settings such as confidence threshold and model size, allowing for a detailed analysis of how each model performs under different conditions. This application is particularly useful for AI researchers and computer vision developers who need to benchmark and understand the nuances of these models in a practical, visual environment.

Microsoft AI Lab (ailab) is a platform designed to empower developers to explore and engage with the latest breakthroughs in Microsoft AI. It offers a unique opportunity to experience, learn, and code with cutting-edge AI technologies. The platform currently features eight distinct projects, demonstrating advancements in areas such as custom vision, attnGAN, Visual Studio tools for AI, Cognitive Search, and Machine Reading Comprehension. Each project provides an experimentation playground, access to source code on GitHub, developer-friendly video tutorials, and insights into the underlying challenges and solutions. Developed in collaboration with Microsoft’s AI School and Microsoft Research (MSR) AI organization, ailab serves as a valuable resource for developers looking to deepen their understanding and practical application of AI.

splat offers a WebGL-based real-time renderer specifically designed for 3D Gaussian Splatting, allowing users to create photorealistic and navigable 3D scenes from a collection of images. This tool is engineered for efficient rendering on typical graphics hardware, extending the capabilities of point cloud rendering. It provides a robust solution for developers and designers looking to generate immersive 3D environments with high fidelity, making advanced 3D scene creation more accessible and performant. The underlying technology focuses on optimizing the rendering process to deliver smooth, interactive experiences.

Object-Detection-Metrics is an open-source toolkit designed to provide comprehensive metrics for evaluating object detection algorithms. It addresses the lack of consensus and standardized implementations for these metrics, offering a reliable solution for researchers and developers. The tool includes implementations for popular metrics such as Intersection Over Union (IOU), Precision, Recall, Precision x Recall curve, and Average Precision (AP), including both 11-point and all-point interpolation methods. It simplifies the evaluation process by accepting ground truth and detected bounding boxes without requiring complex file conversions. The implementation has been carefully compared against official versions, ensuring accurate and trustworthy results for benchmarking different approaches.

Ocean is Meta's in-house, open-source framework designed for Computer Vision (CV) and Augmented Reality (AR) applications. Implemented mainly in C/C++, it offers a platform-independent solution for developers working on advanced CV and AR projects. The framework provides the foundational tools and libraries necessary for building sophisticated applications in these domains. It is released under the MIT License, with specific components under the Art Attribution License 1.0, making it accessible for a wide range of development needs. Developers can leverage Ocean to create applications for various platforms including Android, iOS, Linux, macOS, Meta Quest, and Windows, requiring Python 3.8+, CMake 3.26+, and C++20 for building from source.

ccv is a C-based/Cached/Core Computer Vision Library designed with a minimalism inspiration, making it easy to deploy and integrate into server-side environments. It is highly portable and embeddable, running on various platforms including Mac OSX, Linux, FreeBSD, Windows, iPhone, iPad, Android, and Raspberry Pi. The library implements a range of state-of-the-art algorithms, such as an image classifier, frontal face detector, object detectors for pedestrians and cars, text detection, and general object tracking. A key differentiator is its built-in cache mechanism for image preprocessing, which maintains a clean function interface while transparently handling redundant operations. ccv aims to provide high-performance, modern computer vision implementations, bridging the gap between older, battle-tested algorithms and newer, often MATLAB-based approaches.

Luau is a fast, small, safe, and embeddable scripting language derived from Lua, specifically designed with a gradual type system. It maintains backwards compatibility with Lua 5.1 while incorporating features from future Lua releases and expanding its capabilities with type annotations and a state-of-the-art type inference system. Largely implemented from scratch, Luau's runtime is a heavily modified version of Lua 5.1, featuring a completely rewritten interpreter and other performance innovations. It is widely used by Roblox game developers and engineers for application code and editor plugins, and has seen adoption in other games like Alan Wake 2 and Warframe. The language provides command-line tools for REPL, type checking, and linting, and can be integrated into C++ applications.

This repository offers a comprehensive collection of implementations for popular Reinforcement Learning algorithms, primarily using Python 3, OpenAI Gym, and Tensorflow. It serves as an invaluable learning resource, complementing theoretical materials from "Reinforcement Learning: An Introduction (2nd Edition)" and David Silver's Reinforcement Learning Course. Each folder is structured to correspond with specific chapters or topics, providing learning goals, concept summaries, relevant readings, and practical exercises with solutions. The project covers a wide range of algorithms from Dynamic Programming and Monte Carlo methods to Deep Q-Learning and Policy Gradient methods, making it suitable for both students and researchers in the field.

Accelerate Presentation is a powerful tool designed to streamline the process of launching and training PyTorch models. It enables users to deploy their models across various hardware configurations, including CPUs, GPUs, and TPUs, using a single, unified command. This eliminates the need for extensive code modifications, making the setup and configuration process significantly easier. Hosted on Hugging Face Spaces, Accelerate Presentation provides a user-friendly interface for managing and executing training tasks, ensuring accessibility for developers working with PyTorch. Its core value lies in abstracting away the complexities of distributed training environments, allowing developers to focus on model development rather than infrastructure.

Pokemon Showdown is an application hosted on Hugging Face Spaces that enables users to view any web page within a full-screen iframe by providing its URL. While the name suggests a focus on Pokemon battles, the current functionality described is a general web page viewer. The platform leverages Hugging Face's infrastructure, offering various pricing tiers for compute resources, storage, and inference endpoints. Users can access different CPU and GPU hardware options, including Nvidia T4, L4, L40S, A10G, A100, H100, H200, and B200, with hourly pricing. It also provides options for dedicated inference endpoints and data storage, catering to a range of AI development and deployment needs.