Brief-details: Multilingual HuBERT model supporting 147 languages with 94.4M parameters, trained on 90K hours of speech data. Achieves SOTA in language identification tasks.
Brief Details: FLUX.1-Turbo-Alpha is an 8-step distilled LoRA model for text-to-image generation, optimized from FLUX.1-dev with multi-head discriminator training and fast inference capability.
Brief-details: Efficient 1B param vision-language model optimized for edge devices. Combines Quyen-SE-v0.1 LLM with SigLIP vision encoder. Strong performance on VQA tasks.
Brief-details: Highly efficient GGUF quantization of Qwen 14B model with multiple compression options, optimized for various hardware setups and RAM constraints
Brief-details: RAG-Sequence model for knowledge-intensive NLP tasks, combining retrieval and generation. Built by Facebook, optimized for question-answering using wiki_dpr dataset.
Brief-details: Small-scale Chinese language embedding model (24M params) optimized for text similarity and retrieval tasks, part of BGE v1.5 series with improved similarity distribution.
Brief-details: SQL generation model based on Llama-3 (8B params) optimized for PostgreSQL, Redshift and Snowflake queries. Performs on par with frontier models.
Brief Details: VulBERTa-MLP-D2A is a 125M parameter model for detecting security vulnerabilities in source code, achieving 64.71% accuracy using RoBERTa architecture.
Brief Details: BERT large uncased model with 336M params using whole word masking. Trained on BookCorpus & Wikipedia for masked language modeling tasks.
Brief Details: Hebrew metaphor detection model achieving 95.1% accuracy, fine-tuned on heBERT for 20 common verbs. Popular with 29K+ downloads.
Brief-details: MGM-7B is a 7.27B parameter multimodal model focused on HD image understanding and generation, built on LLaMA/Vicuna with extensive vision-language capabilities.
Brief-details: Habana-optimized configuration for running Llama models on Gaudi HPU processors, featuring fused operations and mixed precision training support
Brief-details: BERT base Japanese model with word-level tokenization using Unidic 2.1.2, trained on Wikipedia with whole word masking. Features 12 layers, 768-dim hidden states, 12 attention heads.
Brief Details: Llama-3.1 based 8B parameter uncensored language model with 77.92% IFEval accuracy. Features enhanced compliance and intelligence.
Brief-details: Large-scale depth estimation model using ViT-L/14 architecture. Trained on unlabeled data, offers state-of-the-art depth prediction capabilities with PyTorch integration.
Brief-details: A Helsinki-NLP English-to-Vietnamese translation model with BLEU score 37.2, using transformer-align architecture and SentencePiece tokenization.
Brief-details: OPUS-MT English-to-Hindi translation model by Helsinki-NLP with BLEU score of 16.1, using transformer-align architecture and SentencePiece tokenization.
Brief Details: T5-base model fine-tuned for sentiment span extraction from tweets, capable of identifying specific text segments that convey sentiment.
Brief Details: A fine-tuned version of Qwen2-VL-7B-Instruct optimized for detailed image captioning, featuring 8.29B parameters and BF16 precision for enhanced descriptions.
Brief-details: Large-scale text-to-image diffusion model supporting Chinese and English inputs, developed by Kuaishou Kolors. Excels in visual quality and complex semantic accuracy.
Brief-details: Uncensored 8B parameter Llama-3 variant optimized for compliance and conversational tasks, using GGUF format. Popular with 29K+ downloads. Requires ethical oversight.
