LG AI Research Unveils EXAONE 3.5: Advanced Language Models for Industrial and Academic Applications

LG AI Research Introduces EXAONE 3.5 Models

LG AI Research has unveiled EXAONE 3.5, a suite of instruction-tuned large language models designed to meet a wide range of needs in the field of artificial intelligence. These models, available in three parameter sizes - 2.4 billion, 7.8 billion, and 32 billion - are now accessible for research purposes, aiming to accelerate progress in both academic and industrial applications 1

1

2

2

.

Advanced Capabilities and Features

The EXAONE 3.5 models boast impressive capabilities, including the ability to handle long contexts of up to 32,768 tokens. This feature enables the models to excel in tasks such as information retrieval, document summarization, and answering multi-hop questions 1

1

2

2

. The researchers highlight the models' proficiency in coding and mathematical problem-solving, as well as their bilingual functionality in English and Korean 1

1

.

Model Variants and Applications

The suite offers three distinct variants to cater to different use cases:

1. The 2.4 billion parameter model: Optimized for on-device AI applications 1

   1

   2

   2

   .
2. The 7.8 billion parameter model: Designed for general-purpose use 3

   3

   .
3. The 32 billion parameter model: Tailored for specialized, high-performance applications 3

   3

   .

All three models demonstrate effectiveness in retrieval-augmented generation tasks, making them valuable tools for both research and practical deployment 1

1

2

2

.

Technological Advancements

LG AI Research has incorporated advanced technologies to enhance the models' performance:

1. Retrieval-Augmented Generation (RAG): This technology generates answers based on real-time web search results or uploaded documents, reducing hallucinations and increasing accuracy 3

   3

   .
2. Multi-step Reasoning: Allows the AI to deconstruct user inquiries step-by-step, developing logical reasoning 3

   3

   .

Open-Source Initiative and Ecosystem Development

In a move to foster an open AI research ecosystem, LG AI Research has made all three EXAONE 3.5 models open-source. This decision is expected to accelerate the pace of innovation in the field 3

3

. The company has also published a technical report on the models' performance, demonstrating confidence in their ability to compete on a global scale 3

3

.

Enterprise Application: ChatEXAONE

LG has launched ChatEXAONE, an enterprise AI agent based on the EXAONE 3.5 model, for its employees. This application incorporates information encryption and privacy protection technology, allowing employees to use it within the company's secure environment 3

3

. ChatEXAONE offers features such as 'Deep' for complex, multi-stage analysis, and 'Dive' for targeted information retrieval 3

3

.

Future Developments and Collaborations

LG AI Research plans to expand the capabilities of EXAONE 3.5, including increasing the token limit to 128,000 in the first half of next year 3

3

. The company has also strengthened collaborations with industry leaders such as AWS, Dell Technologies, Google Cloud, and NVIDIA, among others, to expand the ecosystem and create real-world outcomes with EXAONE and other AI technologies 3

3

.

As AI technology becomes increasingly crucial for national competitiveness, LG's development of these models represents a significant contribution to enhancing AI capabilities on a global scale 3

3

.