CodeV is a 7-billion-parameter vision-language model fine-tuned based on Qwen2.5-VL-7B-Instruct. Through two-stage training of supervised fine-tuning (SFT) and reinforcement learning (RL) based on tool-aware policy optimization (TAPO), it aims to achieve reliable and interpretable visual reasoning. It represents visual tools as executable Python code and ensures the consistency between tool usage and question evidence through a reward mechanism, solving the problem of irrelevant tool invocation under high accuracy.

MultimodalTransformersMultiple Languages

This is the GGUF quantized version of the large language model Mistral-Large-3-675B-Instruct-2512 developed by Mistral AI. The original model has 675 billion parameters and is designed for instruction-following tasks. This project uses the llama.cpp tool and combines it with the imatrix calibration dataset to generate more than 20 different precision quantized model files from Q8_0 to IQ1_S, aiming to balance model performance, inference speed, and storage/memory usage, enabling it to run on a wider range of hardware.

Natural Language ProcessingGgufMultiple Languages

Fara-7B is an efficient small language model specially designed by Microsoft for computer usage scenarios. It has only 7 billion parameters and performs excellently in advanced user tasks such as web operations, competing with larger agent systems.

MultimodalGguf

GLM-4.5-Air-REAP-82B-A12B-mxfp4-mlx is a large language model with 82B parameters compressed by the REAP method based on the GLM-4.5-Air model. It uses the MXFP4 quantization format and is optimized for Mac devices, significantly reducing memory usage while maintaining the performance of core tasks.

Natural Language ProcessingMlxEnglish

The quantized GGUF version of Tencent's Hunyuan Video 1.5 model, specifically designed for text-to-video generation tasks, supports 720P HD video generation, and offers multiple quantization precision versions to optimize usage efficiency.

MultimodalGguf

A quantized model optimized for HunyuanVideo-1.5, used in conjunction with the LightX2V framework, which significantly reduces memory usage while maintaining high-quality video generation performance.

MultimodalDiffusers

Hunyuan Video-1.5 is a lightweight and high-performance video generation model. With only 8.3 billion parameters, it can provide top-notch video quality, significantly lowering the usage threshold. It can run smoothly on consumer-grade GPUs, supporting text-to-video and image-to-video generation, enabling every developer and creator to use it easily.

MultimodalDiffusersMultiple Languages

This is the llama.cpp quantized version of the Precog-123B-v1 model, which provides a variety of quantization types to meet the needs of different hardware configurations and usage scenarios. This model is a large language model with 123 billion parameters and can run efficiently on various hardware after optimization.

Natural Language ProcessingGguf

MiniMax-M2 AWQ - INT4 is a quantized version based on the MiniMax-M2 model. It uses INT4 quantization technology to significantly reduce memory usage and improve inference efficiency while ensuring performance. This model performs excellently in coding and agent tasks and has outstanding comprehensive performance.

Natural Language ProcessingTransformers

Gelato-30B-A3B is a state-of-the-art (SOTA) model fine-tuned for GUI computer usage tasks, offering a quantized version to optimize deployment efficiency. This model is specifically designed to understand and process tasks related to graphical user interfaces.

MultimodalGguf

This project is a quantized version of the text-to-image model Nepotism based on the Flux architecture, optimized using SVDQuant technology. It offers two quantization formats, INT4 and FP4, suitable for different generations of GPU hardware respectively, significantly reducing the model size and memory usage while maintaining image generation quality.

MultimodalDiffusersEnglish

Fara-7B is a small language model developed by Microsoft Research, specifically designed for computer usage scenarios. It has only 7 billion parameters and achieves excellent performance among models of the same scale. It can perform computer interaction tasks such as web automation and multimodal understanding.

MultimodalTransformersEnglish

This project is the result of quantizing the Qwen3-VL-32B-Thinking model. With quantization technology, the model can run more efficiently under different hardware conditions to meet diverse usage needs.

MultimodalGguf

This is a Modal Balance Quantization (MBQ) variant of DeepSeek-OCR. It reduces memory usage by applying 4-bit NF4 quantization to the visual encoder while maintaining the BF16 precision of the projector and language/decoder. It can be locally deployed on ordinary GPUs and used for rapid CPU experiments.

MultimodalTransformers

MiniMax-M2-AWQ is a quantized version based on the MiniMaxAI/MiniMax-M2 model, achieving efficient text generation through the vLLM framework. This model uses the AWQ quantization technology, significantly reducing memory usage and improving inference speed while maintaining model performance. It supports a 32K context length and tool call functionality.

Natural Language ProcessingTransformers

DeepSeek-OCR-bnb-4bit-NF4 is a 4-bit NF4 quantized version of deepseek-ai/DeepSeek-OCR, created using bitsandbytes. While maintaining high accuracy, it significantly reduces video memory usage (up to 8GB), making it very suitable for consumer-grade GPUs.

MultimodalSafetensorsOther

The Impresso NER model is a multilingual named entity recognition model specifically designed for historical document processing. Based on the stacked Transformer architecture, it can identify fine-grained and coarse-grained entity types in digitized historical texts, including person names, titles, locations, etc. The model is optimized for OCR noise, spelling variations, and non-standard language usage in historical documents.

Natural Language ProcessingTransformersMultiple Languages

A quantized version of the text-to-image model based on Flux.1, using SVDQ quantization technology, offering two precisions, INT4 and FP4, suitable for users with different GPU configurations. It significantly reduces the model size and memory usage while maintaining image quality.

Computer VisionDiffusersEnglish

Nanbeige4-3B-Thinking is a 3-billion parameter inference model in the fourth-generation Nanbeige large language model family. It achieves advanced inference capabilities by improving data quality and training methods. This model performs excellently in multiple fields such as mathematics, science, creative writing, and tool usage, and supports multi-stage curriculum learning and reinforcement learning training.

Natural Language ProcessingTransformersMultiple Languages

Gelato-30B-A3B is a state-of-the-art foundation model for GUI computer usage tasks. It is trained on the Click-100k dataset and outperforms previous specialized computer foundation models and larger vision-language models in multiple benchmark tests.

MultimodalTransformersEnglish