🚨 New AI Model: LG AI Research Releases EXAONE 3.5

LG AI Research has released bilingual models expertizing in English and Korean based on EXAONE 3.5 as open source following the success of its predecessor, EXAONE 3.0. The research team has expanded the EXAONE 3.5 models, including three types designed for specific use cases:

✅ The 2.4B model is an ultra-lightweight version optimized for on-device use. It can operate on low-spec GPUs and in environments with limited infrastructure. 

✅ A lightweight 7.8B model offers improved performance over its predecessor, the EXAONE-3.0-7.8B-Instruct model while maintaining versatility for general-purpose use. 

✅ The 32B model represents a frontier-level high-performance option for demanding applications, catering to users who prioritize computational power.

The EXAONE 3.5 models demonstrate exceptional performance and cost-efficiency, achieved through LG AI Research’s innovative R&D methodologies. The hallmark feature of EXAONE 3.5 is its support for long-context processing, allowing the handling of up to 32,768 tokens. This capability makes it effective in addressing the demands of real-world use cases and Retrieval-Augmented Generation (RAG) scenarios, where extended textual inputs are common. Each model in the EXAONE 3.5 series has demonstrated state-of-the-art performance in real-world applications and tasks requiring long-context understanding.leveraging Haystack and deepset to stay ahead in the Gen AI space.

Benchmark Evaluations: Unparalleled Performance of EXAONE 3.5 Bilingual Models

The evaluation benchmarks of EXAONE 3.5 Models were categorized into three groups: real-world use cases, long-context processing, and general domain tasks. Real-world benchmarks evaluated the models’ ability to understand and respond to user queries in practical scenarios. Long-context benchmarks assessed the models’ capability to process and retrieve information from extended textual inputs, which is critical for RAG applications. General domain benchmarks tested the models’ proficiency in mathematics, coding, and knowledge-based tasks. EXAONE 3.5 models consistently performed well across all benchmark categories. The 32B and 7.8B models excelled in real-world use cases and long-context scenarios, often surpassing baseline models of similar size. For example, the 32B model achieved an average score of 74.3 in real-world use cases, significantly outperforming competitors like Qwen 2.5 32B and Gemma 2 27B.

Similarly, in long-context benchmarks, the models demonstrated a superior ability to process and understand extended contexts in both English and Korean. On tests like Needle-in-a-Haystack (NIAH), all three models achieved near-perfect retrieval accuracy, showcasing their robust performance in tasks requiring detailed context comprehension. The 2.4B model was an efficient option for resource-constrained environments, outperforming baseline models of similar size in all categories. Despite its smaller size, it delivered competitive results in general domain tasks, such as solving mathematical problems and writing source code. For example, the 2.4B model scored an average of 63.3 across nine benchmarks in general scenarios, surpassing larger models like Gemma 2 9B in multiple metrics. Real-world use case evaluations incorporated benchmarks like MT-Bench, KoMT-Bench, and LogicKor, where EXAONE 3.5 models were judged on multi-turn responses. They achieved high scores in both English and Korean, highlighting their bilingual proficiency. For instance, the 32B model achieved top-tier results in MT-Bench with a score of 8.51, generating accurate and contextually relevant responses….