Brief Details: State-of-the-art Arabic OCR model based on Qwen2-VL-2B-Instruct, achieving 0.221 WER and supporting diacritics across multiple fonts/layouts.
Brief Details: Powerful 70B parameter Text2SQL model achieving 68.53% accuracy on BIRD test set, using novel ExCoT framework combining Chain-of-Thought with execution-based optimization
BRIEF DETAILS: Text-to-SQL model achieving 68.19% execution accuracy on BIRD test set, using ExCoT framework combining Chain-of-Thought with execution-based DPO optimization.
BRIEF-DETAILS: Gemma 3B quantized model from Google, optimized for efficiency using q4_0 quantization, requiring license acceptance on Hugging Face
Brief-details: State-of-the-art retrieval model built on GTE-Qwen2-7B-instruct, optimized for Chinese/English search. #1 on AIR-Bench with 7B params & 3584d embeddings.
Brief-details: Lumina-mGPT-7B-768 is a 7B parameter multimodal AI model specializing in photorealistic image generation from text, combining vision and language capabilities
Brief-details: OpenThinker2-7B is a fine-tuned version of Qwen2.5-7B-Instruct, trained on OpenThoughts2-1M dataset, achieving state-of-the-art performance in mathematical reasoning and problem-solving tasks.
Brief Details: Vietnamese Text-to-Speech model based on Spark-TTS architecture, offering 0.5B parameters for natural voice synthesis with LLM capabilities
Brief Details: GeometryCrafter is a state-of-the-art model for estimating consistent geometry from open-world videos, running at 1.27-2.49 FPS with advanced point map generation capabilities.
BRIEF-DETAILS: A 7B parameter generative reward model designed to evaluate response accuracy across languages, built on Qwen2.5 architecture for reinforcement learning verification.
Brief-details: HallOumi-8B-classifier is a state-of-the-art 8B parameter model for detecting AI hallucinations, achieving 77.2% macro F1 score and outperforming larger models.
Brief-details: Python-specific AI model from Jane Street with version compatibility requirements - use model_3_11.pt for Python 3.11+ or model.pt for earlier versions
BRIEF DETAILS: A Filipino speech recognition model fine-tuned on facebook/wav2vec2-xls-r-300m, achieving 29.22% WER with linear learning rate scheduling and 30 epochs of training.
Brief-details: AudioX is a versatile Diffusion Transformer model for converting various inputs (text, video, image, audio) into high-quality audio and music, developed by HKUSTAudio.
Brief-details: A compact 1.5B parameter language model optimized for software development tasks, derived from OpenHands LM's larger 32B variant, designed for speculative decoding and local deployment.
BRIEF-DETAILS: 32B parameter coding-focused LLM with 32k context, fine-tuned on Qwen2.5. Optimized for software development with GGUF quantization.
BRIEF-DETAILS: High-performance 32B parameter LLM with multiple quantization options (Q2-Q8) for various hardware configs, optimized with GGUF format and imatrix calibration
Brief Details: A Ghibli-style LoRA fine-tuning of FLUX.1-dev model, optimized for generating Studio Ghibli-inspired artistic images with specific aesthetic qualities
Brief Details: A LoRA model trained on 112 GPT-generated images to create Ghibli-style cartoon art, featuring 64 network dimensions and optimized for FLUX.1-dev base model.
Brief-details: VACE-Annotators: A preprocessing component of the VACE framework for all-in-one video creation and editing, supporting tasks like depth estimation and inpainting.
BRIEF DETAILS: All-in-one video creation and editing model supporting reference-to-video, video-to-video editing, and masked editing with 512x768 resolution capability under RAIL-M license.
