Brief-details: A compact BERT variant that's 7.5x smaller and 9.4x faster than BERT-base, using transformer distillation for efficient NLP tasks.
Brief Details: RoBERTa-based toxicity classifier trained on Jigsaw datasets. Achieves 0.98 AUC-ROC for toxic content detection. Popular with 128K+ downloads.
BRIEF DETAILS: Quantized GGUF variant of Meraj-Mini with 7.62B parameters, offering multiple precision options (2-8 bit) for efficient text generation and conversation tasks.
Brief Details: Quantized version of Mistral-Large-Instruct model with 123B parameters, available in multiple GGUF precision formats (2-8 bit) for efficient deployment.
Brief Details: HuatuoGPT-Vision-7B is a 7.94B parameter multimodal LLM specialized for medical image analysis, built on Qwen2-7B using LLaVA architecture.
Brief-details: FLUX.1-dev-gguf is a quantized text-to-image generation model with 11.9B parameters, optimized for GGUF format and compatible with ComfyUI framework
Brief-details: A compact 22.6M parameter text embedding model optimized for retrieval tasks, achieving SOTA performance with 384-dim embeddings and NDCG@10 of 50.15
Brief-details: Russian speech recognition model based on Whisper Large V3, fine-tuned on Russian datasets with 1.54B parameters, achieving ~10% WER without punctuation.
Brief Details: Inception-v3 model for image classification with 23.9M params, trained on ImageNet-1k. Efficient architecture for computer vision tasks at 299x299 resolution.
Brief Details: NVIDIA's 51B parameter LLM optimized for efficiency, based on Llama-3.1. Features Neural Architecture Search for better performance-to-cost ratio.
Brief-details: Quantized ONNX version of BGE-small for efficient text embeddings and similarity search. Optimized for production with Apache 2.0 license.
Brief-details: E2-TTS is a non-autoregressive zero-shot text-to-speech model trained on the Emilia dataset, offering efficient and high-quality speech synthesis capabilities.
Brief-details: IDM-VTON is an advanced virtual try-on AI model based on SDXL, enabling realistic clothing transfer onto person images with improved diffusion techniques and authentic results.
Brief-details: BERT implementation with Flash-Attention optimization - features configurable attention windows, MLPs, and checkpointing for improved GPU performance
Brief-details: Vietnamese Named Entity Recognition model based on ELECTRA architecture. Achieves 92.14% F1 score on VLSP 2018. Optimized for location, person, organization detection.
Brief-details: PEGASUS-based paraphrasing model fine-tuned for text reformulation with support for multiple output variations. Popular with 132K+ downloads.
Brief-details: Qwen2-VL is a 7B-parameter vision-language model optimized with GPTQ Int4 quantization, featuring dynamic resolution handling and multilingual support for images/videos
Brief Details: Multilingual DeBERTa-v3 model fine-tuned on SQuAD2.0 for question-answering, supporting 94 languages with 278M parameters. Achieves 84% F1 score.
Brief-details: RobeCzech is a 126M-parameter Czech language model based on RoBERTa architecture, trained for masked language modeling with strong performance in NLP tasks.
Brief-details: Vision Transformer model with 304M params, trained on LVD-142M dataset using DINOv2 self-supervised learning. Optimized for image feature extraction.
Brief-details: Named Entity Recognition model for Spanish/English tweets, based on RoBERTuito. Achieves 68.5% accuracy on LinCE benchmark. Popular with 135K+ downloads.
