Brief-details: AutoBench 1.0 is an automated LLM benchmark system using collective LLM-as-a-Judge approach for dynamic, cost-effective model evaluation with high correlation to established benchmarks
Brief-details: SVDQ-FP4-Flux is a quantized model developed by MIT-HAN-Lab, focusing on efficient model compression using FP4 precision and SVD techniques for advanced flux computation.
Brief-details: Flux1-Dev-FP8 is an experimental AI model developed by Academia-SD, focusing on 8-bit floating-point optimization for improved efficiency in deep learning applications.
Brief Details: Experimental FP8 depth-focused AI model developed by Academia-SD, designed for advanced depth perception and processing tasks with optimized memory usage.
Brief Details: Al-Atlas-LLM-0.5B is Morocco's first dedicated 0.5B parameter LLM for Darija dialect, trained on 155M tokens of authentic content.
Brief-details: Experimental 8B parameter LLM focused on roleplay interactions. Features English/Chinese support and specialized playground implementation.
Brief Details: ViT-based deepfake detection model with 99% accuracy. Efficiently distinguishes between real and AI-generated images using patch-based analysis.
Brief-details: Vision Transformer-based deepfake detection model achieving 95.16% accuracy, optimized for 224x224 images. Distinguishes real vs AI-generated images.
Brief-details: COMET-partial is a specialized model for evaluating incomplete translations, built on COMET-early-exit framework, enabling prefix scoring in machine translation quality assessment.
Brief Details: COMET model variant providing layer-wise translation quality estimation with self-confidence scoring, offering both prediction and confidence measures across 25 layers
Brief-details: COMET-instant-confidence is a translation quality estimation model that provides both quality scores and confidence metrics, based on COMET-early-exit architecture
Brief-details: TokenOCR is a pioneering token-level text image foundation model for document understanding, trained on 20M images and 1.8B token-mask pairs, supporting English and Chinese text interaction.
Brief-details: Cybersecurity-focused LLM based on Llama-3.1-8B-Instruct, achieving 15.88% improvement across security benchmarks through specialized training
Brief Details: Japanese BERT variant combining local & global attention, 37M params, 8K context length, trained on 4.39T tokens. Optimized for masked language modeling & fine-tuning.
BRIEF DETAILS: A 70M parameter Japanese BERT variant combining local-global attention, trained on 4.39T tokens with 8K context length. Optimized for masked language modeling and downstream tasks.
Brief Details: LLama3-based 24B parameter model optimized for narrative generation and roleplay. Known for JLLM-style outputs with improved character handling and minimal user impersonation.
BRIEF-DETAILS: BioMedGPT-R1 is a 17B parameter multimodal biomedical reasoning model combining DeepSeek-R1-Distill-Qwen-14B with two-stage training for enhanced medical QA capabilities
Brief-details: Facebook's Data2Vec audio model for self-supervised learning on 16kHz speech, using transformer architecture for contextualized predictions
Brief Details: EXAONE-3.5-7.8B-Instruct is a bilingual (EN/KO) LLM with 6.98B parameters, 32K context window, and state-of-the-art performance in real-world tasks.
Brief Details: Zero123Plus v1.2 is an advanced text-to-3D synthesis model by sudo-ai, building on Zero-1-to-3 technology for enhanced 3D view generation.
Brief-details: CANINE-based model specialized for WTP-split operations, featuring 12 layers without adapters. Focused on efficient text processing tasks.
