Brief-details: DeepSeek-R1-Distill-Qwen-14B GGUF quantized model offering various compression levels from 3.7GB to 12.2GB, optimized for different performance/size tradeoffs
Brief-details: Weighted/imatrix quantized version of oh-dcft-v3.1-gemini-1.5-pro with multiple GGUF variants optimized for different size/quality tradeoffs
BRIEF DETAILS: A 24B parameter Mistral-based model optimized for reasoning, coding, and general tasks. Features GGUF format for local deployment and steerable alignment.
Brief Details: A quantized 4-bit version of Mixtral created by TitanML, optimized for efficient deployment while maintaining performance using AWQ compression technique.
Brief-details: Vision-language model (8B params) built on DeepSeek-R1-Distill-Llama, excelling in VQA and math reasoning tasks with strong performance on benchmarks.
BRIEF DETAILS: A 70B parameter LLaMA-based model combining storytelling, scene description, and reasoning capabilities with unique weight interactions through lorablation and DeepSeek-R1 architecture.
Brief-details: MEMO is an advanced AI model for generating expressive talking videos from single images and audio, using memory-guided diffusion techniques for realistic facial animations.
Brief Details: Prithvi WxC is a 2.3B parameter AI foundation model for weather and climate developed by NASA & IBM, trained on MERRA-2 data for atmospheric state prediction.
Brief-details: ONNX-compatible DPT model with DINOv2-small backbone trained on KITTI dataset for depth estimation, optimized for web deployment via Transformers.js
Brief-details: SDXL implementation of Perturbed-Attention Guidance (PAG) enabling enhanced control over attention mechanisms in image generation with customizable guidance scales
BRIEF DETAILS: Sentence transformer model that creates 768-dimensional embeddings optimized for writing style analysis, with content-independent representations based on authorship verification.
Brief Details: A fine-tuned phoneme recognition model based on wav2vec2-xls-r-300m, trained on TIMIT dataset achieving 7.996% error rate on test set.
BRIEF DETAILS: Vietnamese speech recognition model based on Wav2vec 2.0, fine-tuned on 160 hours of speech data achieving 10.78% WER on Common Voice, without language model.
Brief-details: BERT-based Chinese Named Entity Recognition model achieving 95.25% F1 score. Identifies entities like person names, locations, organizations and time expressions in Chinese text.
Brief Details: A hybrid CNN-Transformer model for document image binarization, converting color/grayscale documents to black & white for improved OCR performance. Developed by SBB.
BRIEF DETAILS: PubMedBERT-based model for biomedical document section classification. Achieves 85.7% test accuracy for labeling abstract sections like BACKGROUND, METHODS, RESULTS.
Brief-details: Swedish NLP model optimized for CPU with tok2vec, tagger, morphologizer, parser, lemmatizer, and NER components. 94.9% lemma accuracy.
Brief-details: XLM-RoBERTa-based NER model trained on HiNER-original dataset, optimized with Adam optimizer and linear learning rate scheduler over 10 epochs
Brief-details: A multilingual word alignment tool based on mBERT that extracts and fine-tunes word alignments between different languages, particularly useful for parallel corpora analysis.
Brief Details: Portuguese doc2query model based on mT5 for document expansion and query generation. Generates multiple queries from text to improve search relevance and training data generation.
Brief Details: Wav2vec2-large-xlsr-galician is a fine-tuned speech recognition model for the Galician language, achieving 7.12% WER on test data using OpenSLR and Common Voice datasets.
