Brief Details: INT4-quantized version of Gemma 3 27B instruction-tuned model, offering efficient deployment while maintaining high performance across reasoning, STEM, and multilingual tasks.
BRIEF-DETAILS: Specialized 12B parameter variant of Google's Gemma designed for unbiased information retrieval, featuring reduced refusal mechanisms and neutral response protocols
Brief Details: A large Chinese embedding model built on Stella, trained on 100M+ samples with hard negative sampling and LLM data synthesis. Optimized for retrieval, classification, and clustering tasks.
BRIEF-DETAILS: LoRA model for text-to-image generation using Flux architecture. Requires TOK trigger word. Built on Replicate's flux-dev-lora-trainer.
Brief Details: An 8B parameter Llama-based language model fine-tuned for reasoning tasks, developed by SciMaker with focus on Taiwan-specific applications
BRIEF-DETAILS: A LoRA model trained on Replicate's Flux trainer, designed for image generation with diffusers library. Uses TOK as trigger word and requires CUDA support.
BRIEF-DETAILS: Infinity is a groundbreaking bitwise autoregressive model for high-res image generation, offering superior quality vs SD3/SDXL with 0.8s generation time at 1024x1024.
BRIEF-DETAILS: First Azerbaijani-focused LLM (7B params) based on LLaMA, achieving 36.7 BLEU score for EN→AZ translation with enhanced fluency and coherence
BRIEF-DETAILS: A collection of LoRA models for Wan2.1-T2V enhancing video generation with features like aesthetics, speed control, high-res fixes, and extended duration support
Brief-details: Qwen2.5-VL-7B-Instruct is a vision-language model featuring dynamic resolution processing, enhanced visual understanding, and support for long video analysis up to 1+ hour with GGUF quantization options
Brief-details: RoboBrain is a unified brain model for robotic manipulation that combines planning, affordance perception, and trajectory prediction capabilities, published in CVPR 2025.
BRIEF-DETAILS: Advanced zero-shot TTS system with GPT-style architecture, featuring Chinese pronunciation correction and precise pause control. Built on XTTS/Tortoise with enhanced speaker features and BigVGAN2.
Brief-details: Advanced 8B parameter LLM fine-tuned for tool/function calling, achieving SOTA performance on Berkeley Function-Calling Leaderboard rivaling GPT-4
Brief Details: Persian-focused 7B parameter LLM optimized for content generation, translation, and Q&A. Features multilingual support with emphasis on Persian language and cultural context.
Brief-details: EraX-WoW-Turbo-V1.1 is a high-speed multilingual speech recognition model, optimized for Vietnamese and 10 other languages, featuring real-time transcription capabilities and ~12% WER.
Brief Details: GGUF conversion of Wan2.1-Fun-14B-InP model, optimized for ComfyUI integration with 14B parameters, specializing in image generation tasks.
Brief Details: Tessa-T1-14B is a specialized React-focused LLM based on Qwen2.5-Coder, optimized for generating semantic React components with advanced reasoning capabilities.
Brief Details: Vietnamese TTS model fine-tuned on 150h of speech data. Supports high-quality voice synthesis with research-only license. Built on F5-TTS base architecture.
Brief-details: A 1.3B parameter text-to-video generation model supporting multi-resolution training and start/end frame prediction, part of Alibaba's Wan2.1 video generation ecosystem
Brief-details: Distil-Large-v3.5 is a knowledge-distilled version of Whisper-Large-v3, offering 1.5x faster inference while maintaining high accuracy for speech recognition tasks, trained on 98k hours of data.
BRIEF-DETAILS: Ling-Coder-lite is a 16.8B parameter MoE LLM optimized for coding, featuring 2.75B activated parameters and 16K context length
