NC unveils language model 'Llama-VARCO LLM'... Top performance among peers

Tuning Language Model Based on Open Source Model 'Llama3.1'

NCSoft unveiled the open-source model-based tuned large language model (LLM) ‘Llama-VARCO LLM’ on the 26th.

Llama-VARCO LLM is a tuned model that enhances Korean language performance by adding NCSoft’s proprietary technology to the open-source model ‘Llama 3.1.’ It is characterized by being small-scale and high-performance. It recorded the top performance among publicly available models with fewer than 10 billion parameters on the ‘Logickor’ benchmark, which measures the reasoning ability of Korean LLMs.

Its broad task compatibility is also a strength. Llama-VARCO LLM can be utilized in various fields such as ▲chatbots ▲QA (quality assurance) ▲natural language generation. In particular, it provides an optimal solution for companies that require small-scale, high-performance LLMs to support Korean AI services.

NCSoft, the first domestic game company to develop its own language model, has been consistently introducing cost-effective models that meet market demands. In April, it introduced the enhanced performance ‘VARCO LLM 2.0,’ which supports multiple languages, and this month it released the evaluation model ‘VARCO Judge LLM’ to verify LLM performance.

To foster AI ecosystem development and support free research, NCSoft is distributing Llama-VARCO LLM on the AI platform ‘Hugging Face’ and providing it as open source. Going forward, the company plans to continue releasing open-source based models to verify its proprietary AI technology and engage with the research community.