Apertus is a fully open multilingual large language model developed by Swiss AI, offering two parameter scales of 7 billion and 8 billion. This model supports over 1000 languages, uses fully compliant and open training data, and its performance can rival that of closed-source models. Apertus was pre-trained on 15T tokens and adopts a phased curriculum training method, supporting a context length of up to 65,536 tokens.

Natural Language ProcessingTransformersOther

Mira is a text generation model based on the fusion of multiple Gemma 3 27B base models. Through carefully selected training data and specific training methods, it has a unique ability to generate poetic texts. This model performs excellently in role-playing and creative writing, and can generate texts with literary charm according to different system prompts.

Natural Language ProcessingTransformers

WEBGEN DEVSTRAL IMAGES is an AI model focused on web page generation. It can generate single-page web pages using HTML, CSS, JS, and Tailwind technologies. This project is trained based on custom templates and uses the supervised fine-tuning method, training with a dataset generated by GPT-OSS-120B.

Natural Language ProcessingTransformersEnglish

PatchFM is a basic model for univariate time series forecasting based on the Transformer architecture. It uses a block-by-block forecasting method, drawing on the training ideas of large language models to divide the time series into blocks for next-block forecasting. It supports multi-quantile output and uncertainty estimation.

Scientific ComputingSafetensors

A lightweight medical fine-tuned model based on Arcee's AFM-4.5B, optimized using a three-stage training method (SFT→DPO→GRPO-CoT), specifically designed for medical education and research scenarios.

Natural Language ProcessingTransformersMultiple Languages

This is a LoRA model that fine-tunes SD3.5-Medium using the DiffusionNFT method. It adopts an innovative online reinforcement learning paradigm and performs excellently in the text-to-image generation task, significantly improving the generation quality and training efficiency.

Computer VisionDiffusers

This is a vision-language model fine-tuned based on the unsloth/qwen2.5-vl-72b-instruct-bnb-4bit model. It is optimized and trained using the Unsloth and Huggingface TRL libraries, and the training speed is twice as fast as the traditional method.

MultimodalTransformersEnglish

Noemie LoRA is the most realistic character LoRA designed specifically for the Qwen model. It uses a unique multi-layer training method to generate images with consistent facial features and high realism.

Computer VisionDiffusers

FLUX.1 [dev] is a rectified flow transformer with 12 billion parameters, specifically designed for text-to-image generation. It excels in image generation quality, second only to FLUX.1 [pro], with excellent prompt following ability and an efficient training method, providing strong support for scientific research and creation.

MultimodalDiffusersEnglish

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

FLUX.1 Krea [dev] is a rectified flow transformer with 12 billion parameters, specifically designed to generate high-quality images based on text descriptions. It excels in output quality, instruction following, etc., focuses on aesthetic photography, adopts an efficient guided distillation training method, and opens its weights to promote scientific research and creative work.

Computer VisionDiffusersEnglish

ICEdit is an innovative instruction-based image editing method that achieves efficient editing through large-scale diffusion transformers, requiring only 0.5% of training data and 1% of parameter scale to achieve SOTA results.

Computer VisionDiffusersMultiple Languages

ICEdit is an instructional image editing method based on large-scale diffusion transformers, achieving state-of-the-art editing effects with only 0.5% of training data and 1% of parameters.

Computer VisionDiffusersMultiple Languages

A text generation model based on google/gemma-3-12b-it, enhanced with the EZO training method to improve Japanese performance

Natural Language ProcessingTransformersJapanese

Ettin is the first collection of encoder - only and decoder - only models trained using the same data, architecture, and training method, supporting fair comparison across scales.

Natural Language ProcessingPytorchEnglish

Tülu 3 is a leading family of instruction-following models, offering fully open-source data, code, and training methodologies as a comprehensive guide to modern technology. Version 3.1 features improvements in the reinforcement learning phase, delivering enhanced overall performance.

Natural Language ProcessingTransformersEnglish

Hunyuan 7B is an open-source bilingual large language model (Chinese and English) by Tencent, featuring optimized data ratios and training methods for robust performance, making it one of the strongest Chinese 7B Dense models available.

Natural Language ProcessingTransformersEnglish

Eagle2 is a high-performance series of vision-language models focused on enhancing model performance through optimized data strategies and training methods. Eagle2-9B is the large model in this series, achieving a good balance between performance and inference speed.

MultimodalTransformersOther

TabPFN is a Transformer-based foundational model for tabular data. Through prior-data-based learning methods, it achieves strong performance on small-scale tabular regression tasks without requiring task-specific training.

AIM-v2 is an image encoder based on the timm library, utilizing distillation training methods, suitable for image feature extraction tasks.

Computer VisionTransformers