Exaone Ecosystem Expands With New AI Models

On 15 July, LG AI Research—the AI R&D arm of South Korea’s LG Group—unveiled Exaone 4.0, a hybrid reasoning AI model that combines general language processing with the advanced reasoning capabilities introduced through the company’s earlier Exaone Deep model.

LG AI Research says its new model outperforms similar models from Alibaba, Microsoft, and Mistral AI in industry benchmarks for science, math, and coding. However, Exaone 4.0 still falls short of Deepseek’s best model.

However, LG AI Research isn’t chasing the same users as most of the familiar names in AI. Unlike models such as ChatGPT and Gemini, which are primarily designed for the average person, LG AI is targeting business users. “Our primary focus is on the business-to-business (B2B) sector rather than business-to-consumer [for now],” says Honglak Lee, the newly appointed co-head of LG AI Research and former research scientist at Google Brain. LG launched the company in December 2020 as part of the Korean tech giant’s digital transformation strategy.

To that end, LG AI Research has made Exaone 4.0 available for research and academic use on Hugging Face, the global open-source AI platform. The model also now supports Spanish language use, expanding its capabilities beyond its original competencies with Korean and English.

Exaone Ecosystem and Strategic Roadmap

Just a week after debuting Exaone 4.0, LG AI Research made its commitment to a B2B focus clear, by unveiling its broader Exaone ecosystem and strategic roadmap. At the AI Talk 2025 on 22 July, the company revealed several new models.

Among the models are Exaone 4.0 Vision Language, a multimodal AI model that can interpret both text and images, and Exaone Path 2.0, a healthcare-focused model designed to diagnose patient conditions in minutes. There are also several enterprise-specific AI agents: ChatExaone, an agent currently being used internally by LG employees to support corporate workflows; Exaone Data Foundry, a platform to accelerate data generation; and an on-premise, full-stack agent that can be deployed in isolated, secure environments without exposing sensitive data.

LG AI Research says that Exaone 4.0 VL, which will be launched in the near future, edges out Meta’s Llama 4 Scout in performance tests. Additionally, the company says that Data Foundry can do in a single day what typically takes 60 experts three months.

Exaone’s on-premise agent runs on chips developed by FuriosaAI, a South Korea-based startup manufacturing neural processing units (NPUs) tailored for AI workloads. According to the company, FuriosaAI‘s RNGD accelerator delivered inference performance on the Exaone models 2.25 times as fast as competing GPUs.

LG also says the hardware to be more energy-efficient. A single rack powered by RNGD chips can generate up to 3.75 times as many tokens for Exaone models than a traditional GPU rack operating within the same power limits.

Autonomous Agents for Enterprise Security

LG AI Research’s ultimate goal is to equip enterprises with all the core components needed to run autonomous agents securely within their own infrastructure, complete with built-in data generation and business operation features, Lee told IEEE Spectrum.

“We’re not just offering an inference engine,” says Lee. “We aim to provide an end-to-end system that integrates the key functionalities enterprises actually need—so they can immediately plug it into their workflow. Every enterprise has unique operational needs. That’s why we’re designing our solution to be flexible—able to combine and configure different parts based on each customer’s environment.”

Further afield, the company is laying the foundation for physical AI, or AI incorporated into robots. “Physical AI is still in its early stages,” says Lee. “But the core framework—perception, reasoning, and action in a continuous loop—is something we’re actively building toward.”

While the company isn’t yet applying this directly to robots, they are demonstrating the same loop with ChatExaone, or the Nexus Agent, an AI agent designed to assess legal compliance of data sets. Crucial to Nexus is the ability to crawl the Internet. “These agents need to understand web pages, extract relevant insights, and act on them,” says Lee. “That’s why we’re building web agents that can navigate complex information flows and make autonomous decisions.”