Projects sharing features with DeepSeek VL2

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

CogVLM is a multimodal large language model designed to integrate visual and textual data for reasoning about images and generating natural language. It functions as a visual question answering system that analyzes image content to provide detailed descriptions or answer specific questions. The project includes a visual grounding model capable of mapping text descriptions to precise bounding box coordinates within an image. It also features a vision-based automation agent that analyzes screen captures to generate execution plans and interaction coordinates for software interfaces. The system

LLaVA is a multimodal large language model architecture designed to process and interpret both image and text inputs to generate natural language responses. It functions as a research-oriented platform for visual instruction tuning, providing a framework to align language models with human intent through training on diverse datasets of paired images and text queries. The system distinguishes itself through a specialized vision-language training pipeline that connects visual data to language models using projection layers and instruction-based fine-tuning. It supports distributed inference by

LAVIS is a multimodal large language model framework and vision-language model library. It provides tools for training and evaluating models that integrate visual, textual, and audio data, serving as a cross-modal feature extractor and a zero-shot visual reasoning engine. The framework distinguishes itself by using frozen-backbone integration, where pretrained encoders remain non-trainable while lightweight adapter layers are updated. It employs cross-modal feature alignment to map different representations into a shared embedding space and utilizes a modular model wrapper to swap vision and

VisualGLM-6B is a multimodal large language model and vision-language system designed to process and generate text based on combined textual and visual inputs. It functions as a bilingual conversational AI capable of maintaining natural language interactions in both English and Chinese. The project utilizes quantized model weights to reduce memory requirements, enabling the deployment of the neural network on consumer-grade hardware. These compressed parameters allow for lower VRAM usage while maintaining the model's ability to analyze visual content and generate corresponding natural languag

ml-ferret is a multimodal large language model framework and visual reasoning engine designed to reason about images and user interfaces. It functions as a UI grounding model and referring expression comprehension tool that maps natural language descriptions to precise pixel coordinates. The system focuses on high-resolution image analysis to identify and locate specific interface components. It employs multi-resolution image processing and region-aware visual encoding to preserve detail across different aspect ratios, enabling the model to analyze spatial relationships and functional layouts

Describe Anything is a multimodal vision-language framework designed for localized visual analysis and automated dataset annotation. It utilizes a vision-language model to generate detailed, context-aware text descriptions for specific regions within images and videos, triggered by user-defined inputs such as points, boxes, or masks. The system distinguishes itself through its ability to maintain object context across video frames via temporal mask propagation and its support for regional question answering without requiring additional model fine-tuning. It provides an OpenAI-compatible API t

Gemma is a family of open-weights large language models based on a decoder-only transformer architecture. These models are designed for text generation and multi-modal conversations, capable of processing and generating responses based on both textual and visual input sequences. The project provides a fine-tunable AI model that supports weight adjustment and low-rank adaptation to specialize performance for particular tasks. It includes support for quantized weights to reduce memory usage and increase inference speed on limited hardware. The capability surface covers multi-modal AI integrati

VisualGLM-6B is a bilingual multimodal large language model and vision-language model designed for conversational tasks and visual understanding. It functions as a bilingual AI model capable of processing and generating responses in both Chinese and English. The system is a quantized large language model supporting 4-bit and 8-bit precision to reduce memory usage and hardware requirements during local deployment. It is also a parameter-efficient fine-tuning model, allowing for weight adjustments to adapt the system to specific downstream tasks without full retraining. The project covers mult

This project is a computer vision dataset and image annotation repository designed for training and evaluating machine learning models. It provides a large collection of labeled images, serving as an object detection benchmark and a source of pixel-level segmentation data. The repository distinguishes itself as a multimodal visual dataset by pairing images with synchronized voice, text, and mouse traces to support narrative understanding. It further enables the analysis of model fairness through the inclusion of demographic attributes and exhaustive annotations. The dataset covers a broad ra

Donut is an OCR-free document transformer and end-to-end document parser. It functions as a neural network that converts unstructured document images directly into structured data or text without the use of an external optical character recognition engine. The project includes a synthetic document generator to create artificial images and ground-truth labels for training. It employs a transformer model to perform visual question answering and document image classification based on visual layout and text. The system covers several document understanding capabilities, including structured info

Qwen2.5-Omni is an omnichannel multimodal large language model designed to process and generate content across text, audio, vision, and video. It functions as a real-time speech AI, utilizing an end-to-end architecture to maintain synchronous voice conversations with low-latency responses. The project emphasizes efficiency through quantized edge models, allowing for local inference on mobile hardware and resource-constrained devices. It employs 4-bit weight quantization, CPU-based process offloading, and on-demand weight loading to reduce GPU memory requirements. The system integrates specia

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

MiniGPT-4 is a multimodal AI framework and large language model that integrates vision encoders with language models to process and reason about combined image and text inputs. It functions as a vision-language model capable of image-based conversational AI, visual question answering, and multimodal logical reasoning. The project utilizes a pretrained vision-language integration strategy that connects a vision encoder to a language model via a linear projection layer. This approach employs frozen-backbone training to align visual representations with linguistic tokens while keeping the primar

Intel XPU LLM Acceleration Library is a toolkit designed to accelerate large language model inference and finetuning on Intel CPUs, GPUs, and NPUs. It provides a distributed inference engine for scaling models across multiple accelerators, a multimodal model runtime for vision and speech tasks, and a low-bit model quantization tool for converting weights into INT4, FP8, and GGUF formats. The project features a parameter-efficient finetuning framework that enables model adaptation using QLoRA and DPO on Intel hardware. It distinguishes itself by providing specialized optimizations for Intel XP

Moondream is a small-scale vision language model designed to reason across images to generate captions and answer natural language questions. It functions as an edge-optimized system capable of performing visual question answering, image captioning, and object detection. The project distinguishes itself through a lightweight architecture designed for local inference on embedded devices, workstations, and air-gapped hardware. It supports the execution of models on local GPUs and Apple Silicon to ensure data privacy and low latency. The system's capabilities include identifying precise object

Janus is a multimodal large language model and unified framework that integrates visual understanding and image generation within a single neural network. It functions as both a visual understanding model for analyzing images and a text-to-image generator. The system uses a unified transformer backbone and a multimodal latent space to bridge the gap between text and visual data. This architecture employs decoupled visual encoding and cross-modal tokenization to separate the paths for discriminative understanding and generative tasks, representing images as grids of discrete codes. The projec

BLIP is a vision-language model framework that combines contrastive, matching, and language modeling objectives to align images with text. Built on a multimodal encoder-decoder architecture, it supports distributed data-parallel training with cosine learning rate scheduling and sliding-window metric tracking for training stability. The framework provides capabilities for image captioning, visual question answering, and cross-modal retrieval, scoring semantic alignment between images and text through learned embeddings. It includes toolkits for fine-tuning pre-trained models on custom datasets

SmolLM is a project dedicated to the development of small language models. It focuses on training and fine-tuning compact models that maintain high performance while utilizing fewer parameters. The project emphasizes efficient AI inference and on-device text generation, aiming to enable the deployment of lightweight models on edge devices with limited memory and processing power. It utilizes synthetic data generation to produce artificial datasets that improve the reasoning and training of these AI systems. The system supports a variety of optimization and training capabilities, including we

Qwen2.5-VL is an autoregressive multimodal transformer designed to process interleaved sequences of text and visual tokens. It integrates visual feature embeddings into a shared language model space to perform cross-modal reasoning and generate coherent responses or structured layout code. The project distinguishes itself through vision-language-action mapping, allowing it to perceive visual interfaces and translate that perception into actionable commands for operating digital screens and robotic hardware. It employs dynamic-resolution image encoding and temporal-frame video indexing to hand

This project provides a foundational framework and reference implementation for executing causal language modeling and multimodal reasoning on local systems. It includes a set of core components for managing model assets, a fine-tuning framework, and structural definitions required to instantiate transformer-based architectures. The system is distinguished by its ability to process combined text and image inputs through multimodal transformer models for visual reasoning and document analysis. It also supports the deployment of quantized models, reducing memory footprints through low-precision

InternVL is a vision-language model framework that fuses a visual encoder with a large language model to translate image features into textual tokens for reasoning. It provides a system for multimodal inference and dialogue, enabling the processing of images and text to answer questions or generate descriptions. The project is distinguished by its high-resolution image processing, which uses dynamic tiling to maintain detail for images up to 4K resolution, and its chain-of-thought visual reasoning for solving complex mathematical and spatial problems. It also supports temporal frame sampling

DeepCamera is an open-source AI video surveillance and network video recorder platform powered by local vision language models and hardware-accelerated processing. It integrates live feeds from network cameras, webcams, and mobile devices to monitor physical spaces while running local edge vision inference without relying on cloud servers. The platform incorporates privacy-preserving video anonymization that converts raw video frames into abstract depth maps in real time, retaining motion tracking while protecting personal identity. Its modular architecture supports pluggable AI scripts and

Data-Juicer is an open-source framework for cleaning, filtering, deduplicating, and transforming multimodal datasets to prepare them for training large language and vision models. It functions as a distributed data pipeline engine that runs processing jobs across Ray clusters, handling billions of samples with automatic operator fusion and adaptive parallelism. The framework provides a library of operators that leverage large language models for semantic extraction, filtering, and data synthesis within processing pipelines. The project distinguishes itself through a YAML-based data recipe sys

This project is a PyTorch implementation of a text-to-image transformer. It is a generative AI model designed to map discrete text tokens to image pixels using a transformer network to create visual content from textual descriptions. The system utilizes a discrete VAE image encoder to compress visual data into tokens for transformer processing. It supports classifier-free guidance to adjust the influence of text prompts during inference and includes capabilities for ranking generated images based on their similarity to text prompts. The architecture incorporates sparse attention mechanisms a

Lucida is a multimodal AI assistant framework and containerized microservice orchestrator. It provides a platform for building agents that process and integrate speech, vision, and text inputs to perform intelligent tasks, supported by a retrieval-augmented generation system for storing and querying factual data from texts, URLs, and images. The framework features a state-graph workflow engine to route user requests through a sequence of microservices using a predefined state machine. It also includes an extensible plugin interface that allows for the integration of custom functional modules