This model is obtained through distillation training using the high-reasoning-difficulty Gemini 3 Pro preview dataset based on the Qwen3-4B-Thinking-2507 base model. It focuses on improving complex reasoning abilities in the fields of coding and science. Through training with specific datasets, it aims to efficiently transfer the reasoning abilities of large models (such as Gemini 3 Pro) to a smaller-scale model.

Natural Language ProcessingTransformers

Nanbeige 4-3B-Thinking-2511 is the latest enhanced version of the Nanbeige series. Through advanced distillation technology and reinforcement learning optimization, it achieves powerful reasoning ability within a compact 3B parameter scale. This model has achieved state-of-the-art (SOTA) results among models with less than 32B parameters in benchmark tests such as Arena-Hard-V2 and BFCL-V4.

Natural Language ProcessingTransformersMultiple Languages

This model is a knowledge distillation model based on the Qwen3-4B architecture, trained with approximately 54.4 million tokens generated by Gemini 2.5 Flash, aiming to integrate the behavior, inference process, and knowledge of Gemini-2.5 Flash into a single dataset.

Natural Language ProcessingTransformersEnglish

This is a fine-tuned version based on Qwen/Qwen3-8B-Base, trained using the TRL framework and the GKD (On-Policy Distillation) method, specifically designed for text generation tasks. The model learns by self-generating errors, improving the quality and stability of text generation.

Natural Language ProcessingTransformers

This is an experimental Stable Diffusion 1.5 distillation model that uses the v-prediction flow matching method and geometric guided adaptive block weighting technology. It is currently in the research phase, the training is ongoing, and the results have not been verified.

Computer VisionDiffusers

This project distills the inference ability of GPT into the Llama-3.1-8B model through an innovative two-stage training process. First, knowledge distillation and format alignment are performed through supervised fine-tuning, and then reinforcement learning is used to encourage the model to independently explore and optimize inference strategies, focusing on breakthroughs in the field of mathematical reasoning.

Natural Language ProcessingTransformersMultiple Languages

This is a vision Transformer model based on the DINOv3 framework, trained on the LVD-1689M dataset from the DINOv3 ViT-7B model through knowledge distillation technology. This model is specifically designed for image feature encoding and can efficiently extract image feature representations, suitable for various computer vision tasks.

Computer VisionTimm

This is a vision Transformer model based on the DINOv3 architecture, using a small configuration and trained through knowledge distillation on the LVD-1689M dataset. This model is specifically designed for efficient image feature extraction and supports various computer vision tasks such as image classification, feature map extraction, and image embedding.

Computer VisionTimm

This is a Vision Transformer image feature encoder based on the DINOv3 architecture, obtained through knowledge distillation from the 7B-parameter DINOv3 ViT model on the LVD-1689M dataset. This model is specifically designed for image feature extraction tasks and has strong visual representation capabilities.

Computer VisionTimm

MobileLLM-Pro is a 1-billion-parameter efficient device-side language model launched by Meta, optimized for mobile devices, supporting a context length of 128k and providing high-quality inference capabilities. This model is trained using knowledge distillation technology, outperforming models of the same scale in multiple benchmark tests, and supporting nearly lossless 4-bit quantization.

Natural Language ProcessingTransformersEnglish

Qwen-Image-Lightning is a text-to-image generation model based on Qwen/Qwen-Image. It achieves fast and high-quality image generation through distillation and LoRA technology and supports convenient calls using the diffusers library.

MultimodalDiffusersMultiple Languages

mdbr-leaf-ir is a high-performance compact text embedding model developed by MongoDB Research specifically for information retrieval tasks. It is particularly suitable for the retrieval stage of the RAG pipeline. The model uses knowledge distillation technology, supports asymmetric architecture, MRL truncation, and vector quantization, and performs excellently in the BEIR benchmark test.

Natural Language ProcessingSentence-transformersEnglish

FastVideo FastWan2.2-TI2V-5B-FullAttn-Diffusers is a text-to-video generation model built on the diffusers library. It is fine-tuned using a novel sparse distillation strategy, supports 3-step fast inference, and can generate videos with a high-quality resolution of 121×704×1280.

MultimodalDiffusers

FLUX.1 Krea [dev] is a rectified flow transformer with 12 billion parameters, capable of generating high-quality images based on text descriptions. This model focuses on aesthetic photography and excels in image quality and instruction following. It uses guided distillation for efficient training and opens its weights to promote scientific research and artistic creation.

MultimodalDiffusersEnglish

FLUX.1 Krea [dev] is an advanced rectified flow transformer model with 12 billion parameters, specifically designed to generate high-quality images based on text descriptions. This model excels in the field of aesthetic photography, has excellent prompt-following ability, is trained using guided distillation technology, and its weights are open for non-commercial use.

Computer VisionDiffusersEnglish

Cogito v2 is a generative hybrid inference model fine-tuned by instructions. It uses iterative distillation and amplification training methods and performs excellently in coding, STEM, instruction following, etc. It supports over 30 languages and a 128k context length.

Natural Language ProcessingTransformers

Cogito v2 is a generative large language model based on instruction tuning, with hybrid reasoning ability, supporting a 128k context length and multilingual processing. This model adopts the Iterative Distillation and Amplification (IDA) training strategy and performs excellently in tasks such as coding, STEM, and instruction following.

Natural Language ProcessingTransformers

Cogito v2 is an instruction-tuned generative model based on Llama-3.1-70B. It has hybrid inference capabilities, supports multiple languages and long contexts, and performs excellently in areas such as coding and STEM. It adopts an iterative distillation and amplification training strategy and can be used for commercial purposes.

Natural Language ProcessingTransformers

FastWan2.1-T2V-1.3B-Diffusers is a video generation model fine-tuned based on the new Sparse-distill strategy. It combines the advantages of distillation and sparse attention, supports efficient 3-step reasoning, and can generate high-quality videos at a resolution of 61×448×832.

MultimodalDiffusers

This is a SPLADE sparse retrieval model based on BERT-Mini (11 million parameters), trained through knowledge distillation of a cross-encoder on the MSMARCO dataset. This model is 6 times smaller than the official splade-v3-distilbert but can achieve 85% of its performance and can process thousands of documents on a CPU.

Natural Language ProcessingSentence-transformersEnglish