Brief-details: SegFormer B1 model fine-tuned for urban scene segmentation on Cityscapes dataset. Features hierarchical Transformer encoder and MLP decode head for efficient semantic segmentation.
Brief-details: CodeLlama-7B-Instruct-GPTQ is a 7B parameter GPTQ-quantized code generation model optimized for instruction-following and coding tasks, offering multiple quantization options for efficient deployment.
Brief Details: Advanced 12B parameter bilingual LLM optimized for Russian/English, featuring RAG capabilities and 128k context, built on Mistral-Nemo architecture
Brief Details: XCiT (Cross-Covariance Image Transformer) image classification model with 12.1M parameters, optimized for 384x384 images with distillation training on ImageNet-1k.
Brief-details: ConvNeXt-Base CLIP model trained on LAION Aesthetic dataset, optimized for 320x320 resolution with augmented regularization, achieving 71.3% ImageNet zero-shot accuracy.
Brief-details: Theia is a vision foundation model for robotics that distills knowledge from multiple vision models, offering 188M parameters with F32 precision and superior performance in robot learning tasks.
Brief Details: A powerful multilingual reranker supporting 75 languages with 306M parameters, featuring SOTA performance and 8192 token context length. Optimized for fast inference.
Brief Details: Korean OCR model with 54.5M parameters, trained on 6M synthetic images. Uses vision-encoder-decoder architecture with DeiT and RoBERTa weights.
Brief Details: VILA1.5-8B is an advanced visual language model built on Llama 3, supporting multi-image reasoning and text generation with 8B parameters.
BRIEF DETAILS: DRAGON+ is a BERT-based dense retriever for efficient text search, featuring dual encoders and strong performance on MARCO Dev (39.0) and BEIR (47.4) benchmarks.
Brief-details: Marigold is a state-of-the-art monocular depth estimation model that repurposes Stable Diffusion for zero-shot depth prediction from single images.
Brief Details: A powerful multilingual Mixtral-8x7B instruction-tuned model optimized for German language use, featuring DPO alignment and 46.7B parameters.
BRIEF DETAILS: A Czech poetry generation model built on GPT-2, specialized in various rhyme schemas (ABBA, ABAB, AABB, AABCCB) from different time periods. 34.9K downloads.
Brief-details: Shap-E is OpenAI's innovative text-to-3D diffusion model that generates textured meshes and neural radiance fields from text prompts, offering fast 3D asset generation.
Here's the two-part response: Brief Details: Korean-optimized 8B parameter LLaMA-3 instruction model, fine-tuned on 60GB+ text data with enhanced Korean language capabilities and chat functionality.
Brief Details: MetricX-23 QE-XL - Advanced reference-free translation quality evaluation model, part of Google's WMT'23 metrics submission
BRIEF DETAILS: Arabic sentiment analysis BERT model fine-tuned on ASTD, ArSAS, and SemEval datasets. Specializes in MSA, dialectal, and classical Arabic text classification.
Brief Details: SmolLM2-360M-Instruct is a compact 362M parameter language model optimized for instruction following, trained on 4T tokens with improved reasoning capabilities.
Brief Details: A powerful 33B parameter code generation model trained on 2T tokens (87% code, 13% language), offering state-of-the-art performance for multiple programming languages.
Brief-details: A 24.9B parameter MoE model available in multiple GGUF quantizations, optimized for English conversation with sizes ranging from 9.4GB to 26.6GB.
Brief-details: A powerful 435M parameter reranking model optimized for search relevance, achieving 48.8 NDCG@10 on BEIR benchmarks with cross-encoder architecture
