Brief-details: Databricks' 2.8B parameter instruction-tuned LLM based on Pythia-2.8b, trained on 15k instruction samples, MIT-licensed for commercial use.
BRIEF DETAILS: ConvNeXt large MLP model (200M params) pretrained on LAION-2B, fine-tuned on ImageNet-12K/1K. Optimized for 320x320 images with strong classification performance.
Brief Details: DenseNet201 is a powerful CNN with 20.2M params, trained on ImageNet-1k. Features dense connectivity pattern and efficient feature reuse.
Brief Details: A 400M parameter conversational AI model by Facebook, trained for open-domain chatting with blended conversational skills and empathy.
BRIEF DETAILS: RoBERTa-large-based sentiment analysis model fine-tuned on 15 diverse datasets, achieving 93.2% accuracy across various text types including reviews and tweets.
BRIEF DETAILS: Vicuna-7b-v1.3: Research-focused chat assistant, fine-tuned from LLaMA on 125K ShareGPT conversations. Open for non-commercial use.
BRIEF DETAILS: RT-DETR object detection model with ResNet101 backbone, trained on COCO and Objects365. Achieves 56.2% AP with 76M parameters at 74 FPS on T4 GPU.
Brief Details: A merged text-to-image model combining InsaneRealistic and Lunar Diffusion, optimized for realistic and artistic image generation with 52K+ downloads
Brief-details: ChatGLM3-6B-Base is a powerful Chinese-English base model with 6B parameters, offering strong performance in semantics, math, reasoning & code generation. Not aligned for chat.
Brief-details: XCiT (Cross-Covariance Image Transformer) with 189M parameters for ImageNet classification, featuring patch-size 8 and 224x224 input resolution. Created by Facebook Research.
BRIEF-DETAILS: ONNX-optimized Splade model for sparse text embeddings and similarity search, with Apache 2.0 license and 52K+ downloads
Brief-details: A Helsinki-NLP English-to-Urdu translation model using transformer-align architecture, achieving 12.1 BLEU score with SentencePiece tokenization
BRIEF DETAILS: Universal image segmentation model using Swin transformer backbone, trained on ADE20k dataset for semantic, instance & panoptic segmentation tasks with MIT license
Brief-details: StableVSR is a cutting-edge diffusion-based video super-resolution model that enhances perceptual quality through temporally-consistent detail synthesis
BRIEF-DETAILS: A student model distilled from zero-shot classification for emotion detection, built on DistilBERT architecture. Optimized for efficient emotion classification using the GoEmotions dataset.
Brief-details: NesT (Nested Hierarchical Transformer) model with 67.7M parameters for image classification, trained on ImageNet-1k. Features hierarchical architecture and efficient visual understanding.
Brief-details: A 7B parameter code-specialized LLM with strong code generation and reasoning capabilities, supporting up to 128K context length and multiple quantization options.
Brief-details: ResNeXt-101 model with 194M params, trained on Instagram-1B dataset using semi-weakly supervised learning and fine-tuned on ImageNet-1k. Achieves 83.35% top-1 accuracy.
Brief Details: Fine-tuned XLSR-53 large model for Finnish speech recognition, achieving 41.60% WER and 8.23% CER on Common Voice dataset, Apache 2.0 licensed.
Brief-details: PixArt-XL-2-1024-MS is a highly efficient text-to-image transformer model that generates 1024px images using only 0.6B parameters, trained with 90% less compute than SD 1.5.
Brief-details: MixNet-L model trained on ImageNet-1k, featuring mixed depthwise convolutions with 7.38M params. Efficient architecture for image classification.
