skt/A.X-4.0-Light

A.X 4.0 Light: Korean-Optimized LLM

A.X 4.0 Light is a 7.6 billion parameter large language model developed by SK Telecom, released as a lightweight version of the A.X 4.0 family. Based on the open-source Qwen2.5 model, it has undergone extensive training with large-scale Korean datasets to deliver strong performance in Korean-centric applications and enterprise settings.

Key Capabilities and Performance

• Superior Korean Proficiency: Achieves a KMMLU score of 64.15 and a CLIcK score of 68.05, demonstrating deep cultural and contextual understanding in Korean.
• Efficient Token Usage: Designed for cost-effective processing, using approximately 33% fewer tokens than GPT-4o for the same Korean input.
• Long Context Handling: Supports a context length of up to 16,384 tokens, enabling comprehension of moderately lengthy documents and conversations.
• Tool-Use Functionality: Capable of tool-use, as demonstrated by examples for function calling and slot filling with external tools like calculate_discount and get_exchange_rate.

Use Cases

This model is particularly well-suited for:

• Korean Language Applications: Ideal for tasks requiring high proficiency in Korean, including translation, content generation, and nuanced understanding.
• Enterprise Deployment: Its lightweight nature and optimization for business environments make it suitable for integration into various enterprise solutions.
• Cost-Effective Processing: Offers an efficient solution for processing Korean text, reducing operational costs compared to larger, less optimized models.
• Instruction Following and General Tasks: Demonstrates solid performance across general benchmarks like Ko-MT-Bench (79.50) and instruction following (Ko-IFEval 72.99).