Brief-details: A cross-lingual reranking model optimized for RAG applications, supporting English, Chinese, Japanese and Korean, with strong performance in document reranking tasks for retrieval-augmented generation.
Brief Details: Kosmos-2.5 is Microsoft's 1.37B parameter multimodal literate model for processing text-rich images, featuring OCR and markdown generation capabilities
BRIEF-DETAILS: BioGPT-Large: Microsoft's biomedical GPT model trained on PubMed data. Excels at medical text generation with state-of-the-art performance on relation extraction tasks.
Brief Details: VLM_WebSight_finetuned is an 8.21B parameter vision-language model that converts website screenshots to HTML/CSS code, built on SigLIP and Mistral-7B.
Brief Details: A quantized version of Falcon 40B optimized for long sequences (11K tokens), using 4-bit GPTQ quantization and dynamic NTK RotaryEmbedding for efficient processing.
Brief-details: Yi-6B-200K is a powerful 6B parameter bilingual LLM from 01.ai with extended 200K context window, trained on 3T tokens for text generation tasks
BRIEF DETAILS: RWKV-4 14B is a powerful 14B parameter language model trained on The Pile dataset, featuring 40 layers and 5120 embedding dimensions with strong performance on various benchmarks. Apache 2.0 licensed.
Brief Details: Würstchen is a highly efficient text-to-image diffusion model achieving 42x spatial compression with novel two-stage architecture, enabling faster inference and training.
Brief-details: A powerful 8B parameter LLM based on Llama-3.1, offering strong instruction-following capabilities with 80.17% accuracy on IFEval and optimized for efficient deployment.
Brief-details: A 13B parameter GPTQ-quantized LLM fine-tuned on 300k+ instructions, offering GPT-3.5-turbo level performance with long responses and low hallucination rates.
Brief-details: World's first 3B-parameter multimodal LLM built on StableLM, capable of vision tasks with ChatML format support
Brief Details: A 70B parameter instruction-tuned LLM merging Platypus2-70B and Llama-2-70b-instruct, optimized for STEM and logical reasoning tasks.
Brief-details: OPT-66B is Meta AI's large-scale decoder-only language model with 66B parameters, trained on 180B tokens for text generation and NLP tasks.
Brief-details: Llama-2-13B-GGML is a quantized version of Meta's 13B parameter LLM, optimized for CPU+GPU inference using GGML format, offering multiple quantization options from 2-bit to 8-bit precision.
Brief-details: SenseVoiceSmall is a multilingual speech foundation model supporting ASR, emotion recognition, and audio event detection across 4 languages with exceptional inference speed.
Brief-details: Llama-2 variant optimized for Korean language with 7B parameters, featuring expanded vocabulary (46,336 tokens) and Korean corpus pretraining
Brief-details: A powerful 67B parameter chat model trained on 2T tokens, built for both English and Chinese language tasks with commercial usage permitted
Brief Details: Llama-2-7b-chat-hf is Meta's 7B parameter chat-optimized LLM, fine-tuned for dialogue with enhanced safety features and commercial usage rights.
BRIEF-DETAILS: SD-Silicon is a general-purpose text-to-image model series created through automated merging (autoMBW), featuring specialized variants and built-in WD1.3 VAE support.
Brief Details: A 15B parameter code-generation model quantized to 4-bit precision, achieving 57.3 pass@1 on HumanEval, optimized for coding tasks
Brief-details: Whisper-Medusa is a 1.56B parameter speech recognition model optimizing Whisper through speculative decoding, enabling faster inference for English audio transcription.
