Awesome GitHub RepositoriesVisual Token Compression

WeUI is a mobile web UI library and design system consisting of CSS components and HTML templates. It is specifically designed to replicate the visual identity and interface of the WeChat messaging ecosystem, providing a standardized set of components to build responsive mobile web interfaces. The library functions as a stateless component system, utilizing a pure CSS architecture and HTML templates that rely on external JavaScript for interactivity. It employs a BEM-based class naming convention to manage component nesting and prevent style leakage across complex layouts. The framework incl

MiniCPM-o is a multimodal large language model designed to function as a real-time conversational assistant on edge devices. By mapping text, image, video, and audio inputs into a unified latent space, the system enables simultaneous cross-modal reasoning and full-duplex interaction. It is built as an edge-side inference engine, utilizing quantized model weights to maintain high-performance processing on consumer hardware. The system distinguishes itself through its integrated speech synthesis and voice cloning capabilities, which allow for the generation of expressive, personalized vocal out

DeepSeek-OCR is a vision processing framework designed to convert image-based text into machine-readable tokens for large language models. It functions as a document inference pipeline that encodes visual data into compact representations, enabling automated optical character recognition and document analysis workflows. The system distinguishes itself through a high-throughput architecture that utilizes hardware-accelerated batch inference to process large volumes of visual data. It incorporates dynamic resolution scaling to manage the balance between visual detail and token consumption, ensu

This project is a comprehensive framework and toolkit for developing, optimizing, and deploying transformer-based models across multimodal, document intelligence, and natural language processing tasks. It provides a unified neural architecture that processes text, vision, audio, and document layout data through a shared set of weights, enabling researchers and developers to build foundational models that align cross-modal representations. The platform distinguishes itself through advanced training and inference strategies designed for large-scale deep learning. It incorporates specialized mec

Qwen2-VL is a multimodal large language model and vision language model designed to process and reason across text, images, and video content. It functions as a visual reasoning engine and a visual agent framework, capable of interpreting visual data to perform object detection, document parsing, and spatial reasoning. The model is distinguished by its ability to act as a video understanding model, processing hour-long videos with second-level indexing and event recall. It further differentiates itself through a visual agent capability that interacts with software interfaces and robotic hardw

This project is a vision language model framework and vision-to-text pipeline designed for deploying and optimizing models that process both images and text. It provides an on-device inference engine and a vision language model framework to run quantized models locally on mobile and desktop hardware accelerators. The framework features a model quantization toolkit to reduce weight precision for lower memory footprints and increased execution speed on specialized silicon. It also includes an efficient vision encoder utilizing a hybrid encoding system to compress image tokens, which reduces pro

OmniRoute is a unified LLM API gateway that connects multiple AI providers to a single endpoint. Its primary purpose is to simplify the integration of various AI models into tools and agents by translating different provider formats into a standardized API. The project distinguishes itself through a multi-strategy request routing system that optimizes for cost, speed, and availability, including automatic model fallbacks and a circuit-breaker resilience model to isolate provider failures. It employs a local-first security posture, using AES-256-GCM encryption to store API keys and conversatio

DeepSeek-VL2 est un modèle de langage multimodal et un système vision-langage conçu pour analyser des scènes visuelles et générer du texte descriptif. Il fonctionne comme un modèle de réponse aux questions visuelles et de mise en correspondance visuelle, capable d'extraire des informations de documents et de localiser des objets ou régions spécifiques dans des images basées sur des descriptions textuelles. Le projet utilise une architecture de mélange d'experts (mixture-of-experts) pour traiter les entrées combinées d'images et de texte. Il est optimisé pour l'inférence via le pré-remplissage incrémentiel, ce qui réduit les besoins en mémoire GPU sur le matériel. Le modèle couvre l'analyse de données multimodales et la compréhension de documents visuels, incluant l'interprétation de graphiques et de mises en page. Il effectue une inférence visuelle et une mise en correspondance pour faire correspondre les requêtes textuelles avec le contenu visuel correspondant.

LLaVA-NeXT est un framework de modèle de langage multimodal et une boîte à outils d'entraînement conçus pour traiter des séquences entrelacées d'images et de vidéos afin de générer du texte. Il fonctionne comme un modèle de langage visuel qui combine des encodeurs de vision avec des modèles de langage pour effectuer des raisonnements complexes, répondre à des questions et comprendre la vidéo. Le système est capable d'analyser des images haute résolution et des trames vidéo temporelles pour décrire des événements, résumer des actions et raisonner à travers plusieurs entrées visuelles. Il prend en charge l'interprétation de documents et de graphiques, l'analyse de l'environnement spatial et la génération de légendes descriptives pour les images et les vidéos. Le framework inclut des outils pour ajuster les modèles multimodaux via l'optimisation des préférences afin de réduire les hallucinations et améliorer la précision. Il fournit également un serveur d'inférence pour déployer ces capacités en tant que service API via un backend HTTP.

ClawRouter is an AI model router and API gateway designed to classify query complexity and assign prompts to the most efficient model tier. It operates as a multi-model AI proxy that orchestrates traffic between various large language models and AI media generators through a unified interface. The project distinguishes itself by integrating a non-custodial micropayment processor using the x402 protocol. This allows for per-request API access and USDC settlement on Base and Solana chains, replacing static API keys with wallet-based authentication and real-time budget enforcement. The system c

SimpleMem is a persistent memory system for AI assistants designed to maintain context across different user chat sessions. It functions as a memory server and multimodal vector database that stores and retrieves information from text, images, audio, and video. The project features a context compression engine that distills interaction histories into compact units to reduce token consumption. It utilizes a distributed memory orchestrator and worker-thread parallel processing to reduce latency when querying large-scale dialogue datasets. The system implements a hybrid indexing approach combin