LG Unveils ExaOne 3.0: Korea's First Open-Source Bilingual AI Model

LG AI Research Launches EXAONE 3.0: A Bilingual Language Model for Open Research

This model is the pioneering open-access member of the EXAONE family, designed to democratize expert-level AI capabilities. LG AI Research has announced the release of EXAONE 3.0, a 7.8 billion parameter instruction-tuned language model, marking a significant advancement in the field of large language models (LLMs). This is the first open model in the EXAONE family, aimed at democratising access to expert-level artificial intelligence capabilities. The release is intended to foster innovation and collaboration within the AI community by providing a high-performance model for non-commercial research purposes. EXAONE 3.0 has been extensively evaluated across a variety of benchmarks, demonstrating competitive real-world performance and instruction-following capabilities. It excels particularly in Korean, while also achieving strong results in English across general tasks and complex reasoning. The model's architecture is based on a decoder-only transformer with advanced features like rotary position embeddings and grouped query attention, supporting a maximum context length of 4,096 tokens. The model was trained using a diverse dataset, ensuring robust performance in real-world scenarios. It features a bilingual tokeniser designed to optimise performance in both English and Korean, addressing the linguistic complexities of the Korean language. The training process involved extensive pre-training and post-training techniques, including supervised fine-tuning and direct preference optimisation, to enhance the model's instruction-following capabilities. The EXAONE 3.0 model is available on Hugging Face for research purposes, supporting the broader AI community's efforts in developing innovative applications. LG AI Research aims to integrate advanced AI into everyday life, making expert knowledge accessible to a wider audience. The model is expected to contribute significantly to advancements in Expert AI, particularly in bilingual environments. LG AI Research has emphasised the importance of responsible AI development, conducting thorough compliance reviews and ethical assessments to minimise risks associated with data usage. The model has been evaluated for potential social and ethical issues, with measures in place to ensure its safe and ethical deployment.

LG Unveils Korea's First Open Source AI, 'EXAONE 3.0' | Taiwan News | Aug. 7, 2024 13:36

- Available for research purposes, contributing to an AI ecosystem SEOUL, South Korea, Aug. 7, 2024 /PRNewswire/ -- LG AI Research today released its latest AI model, EXAONE 3.0, as open source for research purposes. To contribute to the development of the AI research ecosystem, LG AI Research has decided to open source the 'lightweight model' of EXAONE 3.0, which is the most versatile in terms of performance and cost-efficiency, so that it can be utilized for various research purposes. The company is also preparing for the transition to the 'era of everyday AI' by releasing the open beta version of ChatEXAONE, an enterprise AI agent based on EXAONE 3.0, to LG's employees. EXAONE 3.0 proved global competitiveness with top performance among global open-source AI models LG AI Research published a technical report detailing the model training method and performance evaluation findings of EXAONE 3.0. LG AI Research has been focusing on research and development to apply generative AI to real-world industrial applications from its inception, including the release of EXAONE 1.0 in December 2021 and EXAONE 2.0 in July 2023. The latest iteration, EXAONE 3.0, combines enhanced performance with cost-efficiency. Compared to its predecessor, EXAONE 2.0, EXAONE 3.0 shows excellent results in both performance and economy, with 56% less processing time for inference, 35% less memory usage, and 72% lower operating costs. To solve the problem of power consumption triggered by AI, LG AI Research focused on researching lightweight and optimization technologies and succeeded in reducing the size of the model by 97% while increasing performance compared to EXAONE 1.0. In the technical report of EXAONE 3.0, LG AI Research disclosed both the individual scores of the 25 benchmarks used in the evaluation, including MT-Bench, AlpacaEval-2.0, Arena-Hard, and WildBench, which comprehensively evaluate the real-world use cases performance of AI models, as well as the average score for each area, to enhance the reliability of the results. EXAONE 3.0 ranked global top in 13 benchmark scores, including real-world use cases, coding, and math, demonstrating its global competitiveness when compared to other global open-source AI models of similar size, such as Meta's Llama 3.1 and Google's Gemma 2. EXAONE 3.0, a bilingual model that can learn and understand both Korean and English, also recorded the world's best performance in Korean. To improve the performance of EXAONE 3.0, LG's AI Research used more than 60 million data from patents, software code, math, chemistry, and other specialized fields for training. As part of its strategic expansion, the company aims to incorporate additional fields such as law, biology, medicine, education, and various languages by the year-end, to surpass 100 million data points to bolster the system's performance. LG AI Research have also been working to enhance the safety and reliability of EXAONE 3.0 by conducting red-teaming, which involves intentionally attacking the AI model to verify vulnerabilities in technologies, and then improving them. LG products and services with EXAONE 3.0 will be available in the second half of the year LG AI Research will work with LG affiliates to bring EXAONE 3.0 to products and services in the second half of the year. LG AI Research has designed different model sizes for different applications, from the 'ultra-lightweight model' for on-device AI to the 'lightweight model' for general purpose, to the 'high-performance model' for specialized applications. LG affiliates will fine-tune EXAONE 3.0 with their data and apply it to their businesses, products, and services to accelerate innovation. LG AI Research is also actively discussing global partnerships. "As it is important to create AI that can be used in real-world industries, we plan to strengthen partnerships between LG affiliates and other companies and institutions with EXAONE, which has specialized performance and cost-efficiency," said Bae Kyunghoon, President of LG AI Research. "In particular, we plan to open source this highly versatile and lightweight model so that academia and research institutes can utilize the latest generative AI technology, contributing to the AI research ecosystem and further enhancing AI competitiveness." LG's AI assistant, ChatEXAONE, debuted as an Enterprise AI Agent LG will start the open beta service of ChatEXAONE for its affiliates' employees from the 7th. ChatEXAONE is an enterprise AI agent based on EXAONE 3.0 and provides a variety of functions to enhance work productivity, including real-time web-based Q&A, document and image-based Q&A, coding, and database management. LG employees can experience a change in the way they work by using AI for various tasks, from searching to summarization, translation, data analysis, report writing, and coding. ChatEXAONE adopted 'Retrieval-Augmented Generation' (RAG) technology that utilizes real-time web search results to understand the context of the user's instruction (prompts) and provide answers that reflect the latest information. "Generative AI generates different results depending on the prompts, i.e., how the question is asked," said an official from LG AI Research. "We have also applied a feature that recommends sample questions that fit the user's job so that people who are not familiar with generative AI can use it comfortably." ChatEXAONE also provides features for software developers and data analytics experts. LG AI Research expects that ChatEXAONE can generate queries in 22 programming languages (such as Python, Java, and C++,) and SQL (Structured Query Language) for database management, using only natural language (a language used by humans daily) input. This will help users improve their work productivity. LG AI Research will release an open beta service through the end of the year and regularly update it to reflect the feedback and ideas of employees, and will provide the official service and mobile app according to the readiness of each LG affiliate. For affiliates that need to learn proprietary documents and secure data, the company plans to build separate specialized services, like the case of LG Display, which in June built an AI service that can answer questions after training 300,000 additional proprietary documents. Meanwhile, LG has been accelerating its AI transformation for the past four years after establishing the LG AI Research, an AI think tank for LG Group, in December 2020, and has planned to invest KRW 3.6 trillion in R&D in AI by 2026 for future technologies, and global AI talents. Koo Kwang-mo, Chairman and CEO of LG Group has identified AI as a future business and emphasized bold investment and innovation, leading to tangible results such as increasing the number of AI technology applications at each affiliate's business sites and expanding external partnerships, including production processes, product development, and customer service improvement. About LG Group LG Group is a leading global company representing South Korea, offering innovative products and services across various industries such as electronics, chemicals, telecommunications, and energy. Established in 1947, LG Group has grown into a world-renowned brand through its activities in these diverse fields. The company is committed to continuous research and development, focusing on innovation to enhance the quality of life for its customers. Emphasizing its role as a socially responsible enterprise, LG Group is striving to strengthen its competitiveness in the global market and achieve sustainable growth through its future portfolio in areas like AI, Bio, and Cleantech. The company is dedicated to realizing its vision of being a business that provides value to customers and society, pursuing this mission with unwavering determination. About LG AI Research Launched in December 2020 as the artificial intelligence (AI) research hub of South Korea's LG Group, LG AI Research aims to lead the next epoch of artificial intelligence (AI) to realize a promising future by providing optimal research environments and leveraging state-of-the-art AI technologies. And LG AI Research developed its large-scale AI, EXAONE, a 300 billion parametric multimodal AI model, in 2021. EXAONE, which stands for "Expert AI for Everyone," is a multi-modal large-scale AI model that stands out from its peers due to its ability to process both language and visual data. With one of the world's largest learning data capacities, LG AI Research aims to engineer better business decisions through its state-of-the-art artificial intelligence technologies and its continuous effort on fundamental AI research. For more information, visit https://www.lgresearch.ai/.

LG Unveils Korea's First Open Source AI, 'EXAONE 3.0' By Investing.com

- Available for research purposes, contributing to an AI ecosystem , /PRNewswire/ -- LG AI Research today released its latest AI model, EXAONE 3.0, as open source for research purposes. To contribute to the development of the AI research ecosystem, LG AI Research has decided to open source the 'lightweight model' of EXAONE 3.0, which is the most versatile in terms of performance and cost-efficiency, so that it can be utilized for various research purposes. The company is also preparing for the transition to the 'era of everyday AI' by releasing the open beta version of ChatEXAONE, an enterprise AI agent based on EXAONE 3.0, to LG's employees. EXAONE 3.0 proved global competitiveness with top performance among global open-source AI models LG AI Research published a technical report detailing the model training method and performance evaluation findings of EXAONE 3.0. LG AI Research has been focusing on research and development to apply generative AI to real-world industrial applications from its inception, including the release of EXAONE 1.0 in and EXAONE 2.0 in . The latest iteration, EXAONE 3.0, combines enhanced performance with cost-efficiency. Compared to its predecessor, EXAONE 2.0, EXAONE 3.0 shows excellent results in both performance and economy, with 56% less processing time for inference, 35% less memory usage, and 72% lower operating costs. To solve the problem of power consumption triggered by AI, LG AI Research focused on researching lightweight and optimization technologies and succeeded in reducing the size of the model by 97% while increasing performance compared to EXAONE 1.0. In the technical report of EXAONE 3.0, LG AI Research disclosed both the individual scores of the 25 benchmarks used in the evaluation, including MT-Bench, AlpacaEval-2.0, Arena-Hard, and WildBench, which comprehensively evaluate the real-world use cases performance of AI models, as well as the average score for each area, to enhance the reliability of the results. EXAONE 3.0 ranked global top in 13 benchmark scores, including real-world use cases, coding, and math, demonstrating its global competitiveness when compared to other global open-source AI models of similar size, such as Meta (NASDAQ:META)'s Llama 3.1 and Google (NASDAQ:GOOGL)'s Gemma 2. EXAONE 3.0, a bilingual model that can learn and understand both Korean and English, also recorded the world's best performance in Korean. To improve the performance of EXAONE 3.0, LG's AI Research used more than 60 million data from patents, software code, math, chemistry, and other specialized fields for training. As part of its strategic expansion, the company aims to incorporate additional fields such as law, biology, medicine, education, and various languages by the year-end, to surpass 100 million data points to bolster the system's performance. LG AI Research have also been working to enhance the safety and reliability of EXAONE 3.0 by conducting red-teaming, which involves intentionally attacking the AI model to verify vulnerabilities in technologies, and then improving them. LG products and services with EXAONE 3.0 will be available in the second half of the year LG AI Research will work with LG affiliates to bring EXAONE 3.0 to products and services in the second half of the year. LG AI Research has designed different model sizes for different applications, from the 'ultra-lightweight model' for on-device AI to the 'lightweight model' for general purpose, to the 'high-performance model' for specialized applications. LG affiliates will fine-tune EXAONE 3.0 with their data and apply it to their businesses, products, and services to accelerate innovation. LG AI Research is also actively discussing global partnerships. "As it is important to create AI that can be used in real-world industries, we plan to strengthen partnerships between LG affiliates and other companies and institutions with EXAONE, which has specialized performance and cost-efficiency," said Bae Kyunghoon, President of LG AI Research. "In particular, we plan to open source this highly versatile and lightweight model so that academia and research institutes can utilize the latest generative AI technology, contributing to the AI research ecosystem and further enhancing AI competitiveness." LG's AI assistant, ChatEXAONE, debuted as an Enterprise AI Agent LG will start the open beta service of ChatEXAONE for its affiliates' employees from the 7th. ChatEXAONE is an enterprise AI agent based on EXAONE 3.0 and provides a variety of functions to enhance work productivity, including real-time web-based Q&A, document and image-based Q&A, coding, and database management. LG employees can experience a change in the way they work by using AI for various tasks, from searching to summarization, translation, data analysis, report writing, and coding. ChatEXAONE adopted 'Retrieval-Augmented Generation' (RAG) technology that utilizes real-time web search results to understand the context of the user's instruction (prompts) and provide answers that reflect the latest information. "Generative AI generates different results depending on the prompts, i.e., how the question is asked," said an official from LG AI Research. "We have also applied a feature that recommends sample questions that fit the user's job so that people who are not familiar with generative AI can use it comfortably." ChatEXAONE also provides features for software developers and data analytics experts. LG AI Research expects that ChatEXAONE can generate queries in 22 programming languages (such as Python, Java, and C++,) and SQL (Structured Query Language) for database management, using only natural language (a language used by humans daily) input. This will help users improve their work productivity. LG AI Research will release an open beta service through the end of the year and regularly update it to reflect the feedback and ideas of employees, and will provide the official service and mobile app according to the readiness of each LG affiliate. For affiliates that need to learn proprietary documents and secure data, the company plans to build separate specialized services, like the case of LG Display (NYSE:LPL), which in June built an AI service that can answer questions after training 300,000 additional proprietary documents. Meanwhile, LG has been accelerating its AI transformation for the past four years after establishing the LG AI Research, an AI think tank for LG Group, in , and has planned to invest in R&D in AI by 2026 for future technologies, and global AI talents. -mo, Chairman and CEO of LG Group has identified AI as a future business and emphasized bold investment and innovation, leading to tangible results such as increasing the number of AI technology applications at each affiliate's business sites and expanding external partnerships, including production processes, product development, and customer service improvement. About LG Group LG Group is a leading global company representing , offering innovative products and services across various industries such as electronics, chemicals, telecommunications, and energy. Established in 1947, LG Group has grown into a world-renowned brand through its activities in these diverse fields. The company is committed to continuous research and development, focusing on innovation to enhance the quality of life for its customers. Emphasizing its role as a socially responsible enterprise, LG Group is striving to strengthen its competitiveness in the global market and achieve sustainable growth through its future portfolio in areas like AI, Bio, and Cleantech. The company is dedicated to realizing its vision of being a business that provides value to customers and society, pursuing this mission with unwavering determination. About LG AI Research Launched in as the artificial intelligence (AI) research hub of LG Group, LG AI Research aims to lead the next epoch of artificial intelligence (AI) to realize a promising future by providing optimal research environments and leveraging state-of-the-art AI technologies. And LG AI Research developed its large-scale AI, EXAONE, a 300 billion parametric multimodal AI model, in 2021. EXAONE, which stands for "Expert AI for Everyone," is a multi-modal large-scale AI model that stands out from its peers due to its ability to process both language and visual data. With one of the world's largest learning data capacities, LG AI Research aims to engineer better business decisions through its state-of-the-art artificial intelligence technologies and its continuous effort on fundamental AI research. For more information, visit https://www.lgresearch.ai/.

LG unveils Korea's 1st open-source AI EXAONE 3.0

LG Group Chairman Koo Kwang-mo, right, poses with Jim Keller, CEO of Tenstorrent which designs semiconductors for artificial intelligence services, during their meeting in Silicon Valley, Calif., June 23. Courtesy of LG Corp. Products, services powered by new model to roll out in 2nd halfBy Nam Hyun-woo LG Artificial Intelligence (AI) Research announced, Wednesday, the release of its latest AI model, EXAONE 3.0, as open-source for research purposes. This marks the first time a Korean company has launched an open-source AI system, enabling developers to use the model without needing to obtain permission. The company also introduced the open beta version of ChatEXAONE, an enterprise AI agent based on EXAONE 3.0, for LG Group employees. This tool is designed to assist them with their daily work tasks. Logo of EXAONE/ courtesy of LG Group According to the group, LG AI Research has published a technical report detailing the model training methods and performance evaluation results of EXAONE 3.0. The latest model demonstrates superior performance and cost-efficiency compared to its predecessors, EXAONE 1.0, released in December 2021, and EXAONE 2.0, released in July 2023. Compared to EXAONE 2.0, EXAONE 3.0 required 56 percent less processing time for inference, 35 percent less memory usage and 72 percent lower operating costs. The technical report showed that EXAONE 3.0 ranked top in 13 benchmark scores, including real-world use cases, coding and math, demonstrating its global competitiveness when compared to other global open-source AI models of a similar size, such as Meta's Llama 3.1 and Google's Gemma 2. As a bilingual model capable of learning and understanding both Korean and English, EXAONE 3.0 has also received high praise for its performance in Korean. To enhance the performance of EXAONE 3.0, LG AI Research utilized over 60 million data sets from patents, software code, mathematics, chemistry, and other specialized fields for training. The company plans for EXAONE to learn from more than 100 million data sets in fields such as law, biology, medicine, education, and various languages by the end of the year. This chart from LG Group compares the overall performance of EXAONE 3.0 with other open-source AI models. Courtesy of LG Group In the second half of the year, LG AI Research plans to collaborate with LG affiliates to integrate EXAONE 3.0 into various products and services. Depending on its intended use, EXAONE will be available in different model sizes, ranging from an ultra-lightweight version for on-device AI services to a high-performance model for specialized applications. Along with EXAONE 3.0, LG announced the launch of the open beta service for ChatEXAONE, which is available to LG Group employees. Based on EXAONE 3.0, ChatEXAONE provides a variety of functions to enhance work productivity, including real-time web-based Q&A, document and image-based Q&A, coding support and database management. LG AI Research will offer an open beta service through the end of the year, with regular updates based on employee feedback and suggestions. The official service and mobile app will be rolled out according to the readiness of each LG affiliate. "As it is important to create AI that can be used in real-world industries, we plan to strengthen partnerships between LG affiliates and other companies and institutions with EXAONE," LG AI Research President Bae Kyung-hoon said. "In particular, we plan to release the lightweight model as open-source, allowing research institutes to access and utilize our generative AI model, thereby contributing to the AI research ecosystem."