On 15 July, LG AI Analysis—the AI R&D arm of South Korea’s LG Group—unveiled Exaone 4.0, a hybrid reasoning AI mannequin that mixes basic language processing with the superior reasoning capabilities launched by the corporate’s earlier Exaone Deep mannequin.
LG AI Analysis says its new mannequin outperforms comparable fashions from Alibaba, Microsoft, and Mistral AI in trade benchmarks for science, math, and coding. Nonetheless, Exaone 4.0 nonetheless falls wanting Deepseek’s greatest mannequin.
Nonetheless, LG AI Analysis isn’t chasing the identical customers as a lot of the acquainted names in AI. In contrast to fashions akin to ChatGPT and Gemini, that are primarily designed for the typical particular person, LG AI is focusing on enterprise customers. “Our major focus is on the business-to-business (B2B) sector slightly than business-to-consumer [for now],” says Honglak Lee, the newly appointed co-head of LG AI Analysis and former analysis scientist at Google Mind. LG launched the corporate in December 2020 as a part of the Korean tech large’s digital transformation technique.
To that finish, LG AI Analysis has made Exaone 4.0 obtainable for analysis and tutorial use on Hugging Face, the worldwide open-source AI platform. The mannequin additionally now helps Spanish language use, increasing its capabilities past its unique competencies with Korean and English.
Exaone Ecosystem and Strategic Roadmap
Only a week after debuting Exaone 4.0, LG AI Analysis made its dedication to a B2B focus clear, by unveiling its broader Exaone ecosystem and strategic roadmap. On the AI Speak 2025 on 22 July, the corporate revealed a number of new fashions.
Among the many fashions are Exaone 4.0 Imaginative and prescient Language, a multimodal AI mannequin that may interpret each textual content and pictures, and Exaone Path 2.0, a healthcare-focused mannequin designed to diagnose affected person situations in minutes. There are additionally a number of enterprise-specific AI brokers: ChatExaone, an agent presently getting used internally by LG staff to assist company workflows; Exaone Information Foundry, a platform to speed up knowledge technology; and an on-premise, full-stack agent that may be deployed in remoted, safe environments with out exposing delicate knowledge.
LG AI Analysis says that Exaone 4.0 VL, which can be launched within the close to future, edges out Meta’s Llama 4 Scout in efficiency exams. Moreover, the corporate says that Information Foundry can do in a single day what usually takes 60 specialists three months.
Exaone’s on-premise agent runs on chips developed by FuriosaAI, a South Korea-based startup manufacturing neural processing models (NPUs) tailor-made for AI workloads. Based on the corporate, FuriosaAI‘s RNGD accelerator delivered inference efficiency on the Exaone fashions 2.25 occasions as quick as competing GPUs.
LG additionally says the {hardware} to be extra energy-efficient. A single rack powered by RNGD chips can generate as much as 3.75 occasions as many tokens for Exaone fashions than a standard GPU rack working throughout the similar energy limits.
Autonomous Brokers for Enterprise Safety
LG AI Analysis’s final objective is to equip enterprises with all of the core parts wanted to run autonomous brokers securely inside their very own infrastructure, full with built-in knowledge technology and enterprise operation options, Lee advised IEEE Spectrum.
“We’re not simply providing an inference engine,” says Lee. “We purpose to offer an end-to-end system that integrates the important thing functionalities enterprises really want—to allow them to instantly plug it into their workflow. Each enterprise has distinctive operational wants. That’s why we’re designing our answer to be versatile—in a position to mix and configure completely different elements based mostly on every buyer’s setting.”
Additional afield, the corporate is laying the inspiration for bodily AI, or AI integrated into robots. “Bodily AI continues to be in its early levels,” says Lee. “However the core framework—notion, reasoning, and motion in a steady loop—is one thing we’re actively constructing towards.”
Whereas the corporate isn’t but making use of this on to robots, they’re demonstrating the identical loop with ChatExaone, or the Nexus Agent, an AI agent designed to evaluate authorized compliance of knowledge units. Essential to Nexus is the flexibility to crawl the Web. “These brokers want to know net pages, extract related insights, and act on them,” says Lee. “That’s why we’re constructing net brokers that may navigate complicated data flows and make autonomous choices.”
From Your Web site Articles
Associated Articles Across the Internet