Brief-details: A Thai speech recognition model based on wav2vec2-large-xlsr-53, fine-tuned on Common Voice 7.0 dataset, achieving 0.95% WER with PyThaiNLP tokenization
Brief Details: Korean ELECTRA-based sentiment analysis model with 113M params. Specializes in binary classification of Korean text reviews. Apache 2.0 licensed.
Brief-details: Advanced 3D vision model with 689M parameters using ViT architecture for image matching and geometric reconstruction. Created by Naver with CC BY-NC-SA 4.0 license.
Brief-details: A powerful 8B parameter GGUF-quantized LLM based on Llama-3, specializing in function calling, structured outputs, and general conversation with ChatML format support.
Brief-details: Vision Transformer-based image matting model trained on Composition-1k dataset, featuring efficient architecture for foreground object estimation with Apache 2.0 license
Brief Details: A 454M parameter phi3-based model optimized for text generation with 2-bit quantization, offering efficient conversational capabilities through TGI endpoints.
BRIEF DETAILS: MADLAD-400-3B-MT: A powerful multilingual translation model supporting 419 languages, built on T5 architecture with 2.94B parameters. Apache 2.0 licensed.
Brief Details: A fine-tuned Stable Diffusion 2.1 model optimized for realistic image generation, trained on high-quality PhotoChat dataset with 120 curated image-text pairs.
Brief-details: OPT-6.7B is Meta AI's open-source language model with 6.7B parameters, trained on 180B tokens for text generation and NLP tasks.
Brief Details: Large Vision Transformer model pretrained using MAE method. Specialized in masked autoencoding for scalable vision learning. 75% masking ratio, Apache 2.0 licensed.
BRIEF DETAILS: ResNet-18: A lightweight CNN architecture with 11.7M parameters for image classification, featuring residual connections and trained on ImageNet-1k.
BRIEF DETAILS: T5-based paraphrase generation model (223M params) trained on Google PAWS dataset. Capable of generating multiple diverse paraphrases of English sentences.
BRIEF DETAILS: Multilingual zero-shot text classifier (568M params) based on BGE-M3, capable of processing 8K tokens with strong performance across 28 classification tasks.
Brief-details: Japanese BERT model using character-level tokenization, trained on CC-100 & Wikipedia data with whole word masking. Features 12-layer architecture with 768-dim hidden states.
BRIEF DETAILS: Self-supervised interest point detector & descriptor model with 1.3M params, optimized for keypoint detection and feature extraction in computer vision tasks
BRIEF DETAILS: A lightweight ChatGLM2 variant with 19.3M parameters, optimized for feature extraction using transformer architecture. Uses F32 tensor type and cosine learning rate scheduling.
Brief Details: A 3.32B parameter code completion model trained on 1T tokens across 30 programming languages, optimized for code generation and completion tasks.
Brief-details: Gemma 2 9B IT GGUF is Google's lightweight LLM (9.24B params) optimized for efficiency, featuring multiple GGUF quantization variants for flexible deployment.
Brief-details: A lightweight 0.5B parameter Qwen2.5 instruction-tuned model optimized for MLC-LLM and WebLLM deployment, featuring q4f16_1 quantization for efficient execution.
Brief-details: Vision Transformer (ViT) huge model with 632M params, trained on ImageNet-21k. Specialized in image classification with 14x14 patch size at 224x224 resolution.
Brief-details: LCM-LoRA SDXL: High-speed inference adapter for Stable Diffusion XL, enables 2-8 step generation with 197M parameters. Supports text-to-image, inpainting, and ControlNet.
