Brief Details: A Korean-optimized Whisper tiny model trained using federated learning techniques for speech recognition tasks, focusing on Korean language processing.
Brief-details: A specialized video diffusion model that treats diffusion as a shader process, enabling 3D-aware video generation with advanced control capabilities and versatile editing options.
Brief-details: A specialized LoRA model for FLUX.1 focused on halftone car photography effects, optimized for 1280x832 resolution with 49 training images.
Brief-details: Audio Flamingo 2 is a 3B parameter audio-language model from NVIDIA that achieves SOTA performance in audio understanding and expert reasoning, capable of processing 5-minute audio clips
Brief-details: Akshara-8B is an 8B parameter multilingual LLM optimized for Indian languages, supporting 8 languages including Hindi, Tamil, and English. Built by SVECTOR.
BRIEF DETAILS: 12B parameter GGUF quantized model offering multiple compression variants (Q2-Q8), with optimized versions balancing quality and performance.
Brief-details: Greek-English instruction-tuned 8B parameter LLM built on Llama-3.1, optimized for 128k context, strong bilingual capabilities and domain expertise
BRIEF-DETAILS: 12B parameter language model merging Rei-12B and Francois-Huali-12B, optimized for roleplay and creative writing using ChatML format
Brief-details: Advanced sentence transformer model built on ModernBERT-large, optimized for semantic search with 1024D embeddings and 8192 token support
BRIEF-DETAILS: Arcee-Blitz (24B) - A Mistral-based model distilled from DeepSeek, offering enhanced performance across various tasks with improved world knowledge and efficiency.
Brief-details: Advanced multilingual translation model based on Qwen 2.5 32B, fine-tuned for 16 languages using RLHF and expert feedback. Optimized with LoRA + QLoRA.
BRIEF-DETAILS: 24B parameter Mistral-based model optimized for fine-tuning. Features ChatML support, creative writing capabilities, and roleplay functionality. Intentionally undercooked at loss ~8.0.
Brief-details: q3-reasoner is a fine-tuned version of Qwen2.5-Coder-3B-Instruct optimized for faster inference using Unsloth and TRL, licensed under Apache-2.0
BRIEF DETAILS: Fast tissue segmentation model for H&E pathology slides using UNet with MobileNet-v3 encoder. Achieves 0.93 mIoU, processes slides in <1s on CPU.
Brief-details: QLIP-L-14-392 is NVIDIA's state-of-the-art visual tokenization model combining high-quality image reconstruction with zero-shot image understanding, achieving 79.1% accuracy.
Brief-details: Powerful merged LLM combining Phi-4 variants using SLERP method. Ranks #3 among models up to 15B parameters with strong performance on various benchmarks.
BRIEF-DETAILS: A merged 14B parameter LLM combining Lamarck and Qwenvergence models, ranking #1 among sub-15B models with impressive performance in reasoning tasks
Brief-details: Quantized version of DeepSeek-R1 that maintains full accuracy while reducing size by 75%. Achieves 17.2 tokens/sec with just 5GB RAM usage.
Brief-details: 8-bit quantized version of DeepSeek-R1-Distill-Qwen-7B offering 1.6x speedup and 50% memory reduction while maintaining accuracy
Brief Details: 8-bit quantized version of DeepSeek-R1-Distill-Qwen-32B offering 2x faster inference with 99.57% accuracy retention and 50% memory reduction
BRIEF-DETAILS: Boreal-HL is a specialized LoRA for Hunyuan that enhances detail, depth of field, skin textures and lighting in both video and image generations.
