NewenAI/QuettaLLMs-27B-Koreasoner-V3
NewenAI/QuettaLLMs-27B-Koreasoner-V3 is a 27 billion parameter Korean large language model, fine-tuned by NewenAI from the Qwen/Qwen3.5-27B base model. It specializes in logical reasoning and Korean knowledge, delivering direct and concise answers to optimize token usage and response speed. The model demonstrates improved performance on Korean reasoning and knowledge benchmarks, making it efficient for QA and automated evaluation in Korean contexts. It features a 32768 token context length and is designed to reduce hallucination and improve math problem-solving accuracy.
Loading preview...
QuettaLLMs-27B-Koreasoner-V3 Overview
QuettaLLMs-27B-Koreasoner-V3 is a 27 billion parameter Korean large language model developed by NewenAI, built upon the Qwen/Qwen3.5-27B base. It has been fine-tuned using LoRA to significantly enhance its logical reasoning capabilities and Korean domain understanding.
A key differentiator of this model is its output efficiency. While it performs complex internal reasoning, it is specifically trained to provide direct and concise final answers, minimizing verbose explanations. This design choice reduces token consumption and accelerates response times, making it highly suitable for applications requiring efficient QA and automated evaluation.
Key Capabilities & Features
- Enhanced Korean Reasoning: Improved scores on Korean reasoning and knowledge evaluation metrics, including KMMLU-Pro (+0.054), CLIcK (+0.024), and KMMLU (+5.30p) compared to its base model.
- Output Efficiency: Delivers concise answers, optimizing token usage and response speed.
- Reduced Hallucination: Demonstrates decreased information distortion, particularly in multi-step logical tasks.
- Korean Context Processing: Stronger understanding of Korean background knowledge due to specialized training data.
- Math Problem Solving: Improved accuracy in mathematical calculations through internal reasoning processes.
- Zero-Shot Performance: Consistently provides formatted answers without requiring additional prompt examples.
Training & Data
The model was trained with over 230,000 refined SFT samples, including Korean knowledge data (KMMLU-format), logical reasoning data (public and synthetic), and math/quantitative reasoning datasets. The training focused on guiding the model to identify core questions and generate accurate, concise answers.