Brief Details: Large Vision Transformer model with 304M params, using registers and DINOv2 pretraining on LVD-142M dataset. Optimized for image feature extraction.
Brief-details: Cross-encoder model optimized for MS Marco passage ranking, achieving 74.30 NDCG@10 on TREC DL 19, processes 1800 docs/sec on V100 GPU.
Brief-details: Multilingual text-to-text transformer model supporting 101 languages, pre-trained on mC4 dataset. Large-scale variant ideal for cross-lingual NLP tasks.
Brief-details: A Vision Transformer (ViT) model fine-tuned for uppercase English character recognition, achieving 95.73% accuracy using the google/vit-base-patch16-224-in21k architecture.
Brief Details: ColPali v1.2 - An efficient visual document retriever combining PaliGemma-3B with ColBERT strategy. Enhanced with right padding and improved training parameters.
Brief-details: Meta's 8B parameter instruction-tuned LLM optimized for dialogue. Features enhanced safety, reduced false refusals, and 8k context length. Strong performance on benchmarks.
BRIEF DETAILS: Fine-tuned FlanT5-XXL model specialized for image-text retrieval tasks, developed by Zhiqiu Lin. Apache-2.0 licensed with CLIP integration.
Brief-details: State-of-the-art embedding model ranking #1 on MTEB benchmark, built on Mistral-7B with 4096d embeddings and latent-attention pooling
BRIEF-DETAILS: Helsinki-NLP's Polish-to-English translation model with BLEU score of 54.9, built on Marian architecture using OPUS dataset and transformer-align approach.
BRIEF DETAILS: 70B parameter instruction-tuned Llama 3.1 model with multiple GGUF quantization options (19-75GB), optimized for different performance/size tradeoffs.
BRIEF DETAILS: Meta's Llama-3 70B model optimized with FP8 quantization, reducing memory footprint by 50% while maintaining 99.55% accuracy. Ideal for commercial and research applications.
Brief-details: Creative ControlNet model for generating scannable QR codes with artistic designs. Supports SD-1.5, features gray background integration and adjustable creativity/readability balance.
Brief-details: Language-centric multimodal model that binds video, audio, thermal, depth and image modalities through language alignment, achieving SOTA performance across multiple datasets
BRIEF DETAILS: WangchanBERTa - A 106M parameter RoBERTa-based Thai language model trained on 78.5GB text, specialized in masked language modeling and classification tasks.
Brief Details: GPT-NeoX-20B: 20B parameter open-source language model by EleutherAI, trained on The Pile dataset. Strong performance in zero-shot tasks.
Brief-details: A wide ResNet-50 variant with RandAugment (RACM) training, achieving 82.27% top-1 accuracy on ImageNet, featuring 68.9M parameters and optimized training.
BRIEF-DETAILS: Cutting-edge multilingual translation model supporting 101 languages for speech and 96 for text, with 2.31B parameters and improved UnitY2 architecture
BRIEF DETAILS: BERT-based context encoder for dense passage retrieval, trained on multiple QA datasets including NQ, TriviaQA, WebQuestions, and CuratedTREC. Optimized for open-domain question answering.
Brief Details: Portuguese legal NER model based on BERT, achieves 89.3% F1 score, specialized for legal entity recognition with strong performance on person, organization and temporal entities
Brief-details: Facebook's Wav2Vec2 speech recognition model trained on 960h of LibriSpeech data. Achieves 1.9/3.9 WER on clean/other test sets. Self-training enhanced.
Brief Details: T5-base model fine-tuned for tag generation from article content, featuring 223M parameters and achieving 38.6% ROUGE1 score on evaluation.
