Wan2.2 is a major upgrade version of the basic video model, focusing on integrating innovative technologies such as effective MoE architecture, efficient training strategies, and multimodal fusion into the video diffusion model, bringing more powerful and efficient solutions to the field of video generation.

MultimodalDiffusersMultiple Languages

GigaChat3-10B-A1.8B is a dialogue model in the GigaChat series, based on the Mixture of Experts (MoE) architecture, with a total of 10 billion parameters, of which 1.8 billion are active parameters. This model uses multi-head latent attention and multi-token prediction technologies, supports a long context of 256,000 tokens, and performs excellently in multilingual dialogue and reasoning tasks.

Natural Language ProcessingSafetensorsMultiple Languages

GigaChat3-10B-A1.8B-base is the basic pre-trained model of the GigaChat series, adopting the Mixture of Experts (MoE) architecture with a total of 10 billion parameters and 1.8 billion active parameters. The model integrates Multi-Head Latent Attention (MLA) and Multi-Token Prediction (MTP) technologies, and has the advantage of high throughput during inference.

Natural Language ProcessingSafetensorsMultiple Languages

This is a large language model with 139B parameters obtained by uniformly pruning 40% of the experts in MiniMax-M2 using the REAP method. It adopts the GLM architecture and Mixture of Experts (MoE) technology, and undergoes various quantization processes through llama.cpp, suitable for text generation tasks.

Natural Language ProcessingGgufEnglish

Fyodor-StarCoder2-7B-MoE is an enhanced version based on StarCoder2-7B, adopting a Mixture of Experts architecture with 3 MoE layers, 6 expert networks, and a top-2 routing mechanism. It performs excellently in code generation, function calls, and agent tasks.

Natural Language ProcessingTransformersOther

Moondream 3 Preview HF is a re-implementation of the Moondream 3 (Preview) model based on the HuggingFace Transformers architecture specification, making it fully compatible with the Hugging Face ecosystem. It is a multimodal vision-language model that uses a Mixture of Experts (MoE) text backbone, with approximately 9 billion parameters and 2 billion active parameters.

MultimodalTransformers

Qwen3-VL is the most powerful vision-language model in the Tongyi series. It adopts a Mixture of Experts (MoE) architecture and provides weights in the GGUF format, supporting efficient inference on devices such as CPUs and GPUs. The model has been comprehensively upgraded in text understanding, visual perception, spatial understanding, video processing, etc.

MultimodalGguf

Qwen3-VL is the most powerful vision-language model in the Qwen series to date. It has been comprehensively upgraded in text understanding and generation, visual perception and reasoning, context length, spatial and video dynamic understanding, and agent interaction capabilities. This model adopts a Mixture of Experts (MoE) architecture, providing excellent multimodal processing capabilities.

MultimodalGguf

Qwen3-VL is the most powerful vision-language model in the Tongyi large model series. It has excellent text understanding and generation capabilities, in-depth visual perception and reasoning capabilities, long context support, strong spatial and video dynamic understanding capabilities, and outstanding intelligent agent interaction capabilities. This model uses a Mixture of Experts (MoE) architecture and is a thinking version for enhanced reasoning.

MultimodalGguf

Qwen3-Next-80B-A3B-Thinking is a new-generation thinking large language model launched by the Tongyi Qianwen team. It adopts an innovative hybrid attention mechanism and a highly sparse MoE architecture. While maintaining efficient reasoning, it has powerful complex reasoning capabilities and natively supports a context length of 262K.

Natural Language ProcessingTransformers

This is a collection of quantized versions of the inclusionAI/Ling-1T model, developed based on the ik_llama.cpp branch. It offers multiple quantization schemes to meet different memory and performance requirements. These quantized models provide excellent perplexity performance under a given memory footprint and are specifically optimized for the inference efficiency of the MoE architecture.

Natural Language ProcessingGguf

The Ming-flash-omni preview version is a multimodal large model built on the Ling-Flash-2.0 sparse mixture of experts (MoE) architecture, with a total of 100B parameters and only 6B parameters activated per token. This model is a comprehensive upgrade based on Ming-Omni, showing significant improvements in multimodal understanding and generation, especially in speech recognition, image generation, and segmentation editing.

MultimodalDiffusersEnglish

LFM2-8B-A1B is an 8-bit quantized MLX build version optimized for Apple Silicon chips, adopting the Mixture of Experts (MoE) architecture. It has approximately 8 billion total parameters, with about 1 billion parameters activated per token, and supports fast inference on the device side.

Natural Language ProcessingMlxEnglish

NVIDIA GPT-OSS-120B Eagle3 is an optimized version based on the OpenAI gpt-oss-120b model. It adopts the Mixture of Experts (MoE) architecture, with a total of 120 billion parameters and 5 billion active parameters. This model supports both commercial and non-commercial use and is suitable for text generation tasks, especially for the development of AI Agent systems, chatbots, and other applications.

Natural Language ProcessingSafetensors

Wan2.2 is a significant upgrade version of the basic video model. It introduces the Mixture of Experts (MoE) architecture, incorporates carefully curated aesthetic data, and is trained on larger datasets to enhance the ability to generate complex motions. This model supports generating 5-second videos with 480P and 720P resolutions, and has significant improvements in video generation quality and performance.

MultimodalDiffusers

The 5-bit quantized version of the IBM Granite-4.0-H-Tiny model, optimized for Apple silicon chips. It uses a hybrid architecture of Mamba-2 and soft attention, combined with a Mixture of Experts model (MoE), to achieve efficient inference while maintaining high quality.

Natural Language ProcessingMlxMultiple Languages

IBM Granite-4.0-H-Tiny is a hybrid Mamba-2/Transformer model optimized for Apple silicon chips, using 3-bit quantization technology, and is designed for long context, efficient inference, and enterprise use. This model combines the Mamba-2 architecture with the Mixture of Experts (MoE) technology, significantly reducing memory usage while maintaining expressiveness.

Natural Language ProcessingMlxMultiple Languages

Qwen3-Next-80B-A3B-Instruct is an efficient sparse mixture of experts model developed by the Tongyi Qianwen team. It has a total of 80 billion parameters, with only 3 billion active parameters. The model adopts an innovative hybrid attention mechanism and a MoE architecture with an extremely low activation rate, significantly improving inference efficiency while maintaining strong performance. It natively supports a context length of 262,000 tokens and can be extended to 1 million tokens.

Natural Language ProcessingTransformers

Megrez 2-3x7B-A3B is a device-native large language model that adopts the Mixture of Experts (MoE) architecture, combining the accuracy of MoE and the compactness of dense models. The model is trained on 8T token data, has a context length of 32K, and plans to improve inference and agent capabilities in the future.

Natural Language ProcessingGgufMultiple Languages

Qwen3-Next-80B-A3B-Thinking-AWQ-4bit is a 4-bit version generated from the Qwen3-Next-80B-A3B-Thinking model through the AWQ quantization technology, which significantly improves the model's inference efficiency. This model adopts a hybrid attention mechanism and a highly sparse MoE architecture, and performs excellently in complex reasoning tasks. It supports a context length of up to 262K tokens.

Natural Language ProcessingTransformers