LG AI Unveils EXAONE 3.5 Models for Industrial Use

LG AI launched EXAONE 3.5, a suite of instruction-tuned large language models tailored to meet a wide range of needs, from compact deployments to resource-intensive tasks.  

The models, offered in three parameter sizes—2.4 billion, 7.8 billion, and 32 billion, are available for research purposes. 

According to the study, these models are designed to speed up progress in both academic and industrial applications by providing versatile solutions for generative AI innovation.

Key Capabilities

The research report notes that EXAONE 3.5 models are adept at handling real-world challenges and performing tasks that involve long contexts of up to 32,768 tokens. The researchers further say that the model is proficient in coding and mathematical problem-solving. It supports bilingual functionalities in English and Korean. Benchmark tests highlight their competitive performance.

The model’s features are designed to be efficient, effective, and user-friendly. The AI models are trained on large datasets using advanced methods, enabling them to follow instructions accurately and understand long inputs or contexts. They are optimised to meet user needs effectively while keeping training costs lower than other similar models.

Limitations include occasional inaccuracies or biases stemming from training data and the lack of real-time updates, which can result in outdated information.

Real-world use cases

The EXAONE 3.5 models are equipped to tackle a variety of tasks. For long-context applications, they excel in retrieving information, summarising documents, and answering multi-hop questions.  

In general domains, they solve advanced math problems and address knowledge-based queries. 

The 2.4B variant is optimised for on-device AI, while all models are effective in retrieval-augmented generation tasks, making them invaluable for research and practical deployment.