Summary
LG AI Research
Enterprise AI (LG Group)
EXAONE 3.0, 3.5, 4.0, 4.5 and K-EXAONE 236B-A23B (MoE)
NXT RNGD Server
POWER EFFICIENCY VS. A100
2.3×
EXAONE MODELS VALIDATED
5
Largest model validated, on 4 RNGD cards
236B
Power Efficiency
LG is a global enterprise spanning electronics, telecommunications, IT services and chemicals. LG AI Research leads foundation model research and AI transformation across the group, developing the EXAONE family and applications ranging from ChatEXAONE, an assistant used within LG, to models powering customer at its affiliates.
Supporting this breadth of AI applications requires efficient inference as models and workloads evolve. LG AI Research needed inference compute that could complement its GPU infrastructure, improve cost efficiency per token, and support each new EXAONE generation.
Furiosa’s RNGD has been validated across five EXAONE models: the 32B-class dense models EXAONE 3.0, 3.5, 4.0 and 4.5, and K-EXAONE, a 236B-parameter mixture-of-experts model. In FP8 inference testing, the NXT RNGD Server measured 2.3× the performance per watt of a server powered by A100 SXM GPUs.
Challenge
Enterprise AI Needs Performant and Efficient Serving
Foundation models power a growing range of enterprise AI applications, from internal assistants and chatbots to customer-facing services. At LG’s scale, the cost and efficiency of serving these models can have a meaningful impact on the economics of deploying AI across every business unit. LG AI Research needed inference compute that could complement its existing GPU infrastructure while improving serving efficiency.
Keeping pace with EXAONE’s development presented a second challenge. As LG AI Research released new generations, changes in model architecture, precision requirements, and capabilities meant the inference platform needed to support evolving workloads without slowing deployment.
Solution
One Serving Stack Across EXAONE Generations
Furiosa built a repeatable serving configuration for both dense and MoE EXAONE models. Common configurations for quantization, parallelism, and parsing provide a consistent starting point for each new generation, with model-specific updates delivered through scheduled SDK releases.
From 32B Dense to 236B MoE
EXAONE 3.0 through 4.5 (32B-class dense models) run in FP8 using tensor parallelism across four RNGD cards in a single server. K-EXAONE, a 236B-parameter MoE model, runs in NVFP4A16 with 4-bit weights and 16-bit activations. Its pre-built serving artifact fits within the 192 GB of HBM3 across four Furiosa RNGD cards, with all 32 processing elements in a single tensor-parallel group. The same server configuration supports both ends of the model range.
Built for Agent Workloads and Repeatable Transitions
Server-side reasoning-token separation and tool-call parsing, including automatic tool selection, support agent integrations without requiring client-side changes. Each model update follows a repeatable process: post-training quantization, quality-regression testing against the reference model, and artifact release. EXAONE 4.5-33B completed this process in SDK 2026.4. Prometheus and Grafana provide ongoing visibility into KV-cache usage and prefix-cache hit rates.
Development
From Validation to Production
FP8 EXAONE models are already serving always-on internal services, including ChatEXAONE. K-EXAONE has completed validation and is being prepared for production. With model configurations and precision settings validated in advance, LG AI Research can build on an established serving setup as it expands deployment across the group.
VISION
Supporting the Next Generation of EXAONE
As EXAONE evolves and AI applications expand across LG’s businesses, each new model generation needs a clear path to deployment. LG AI Research and Furiosa are working to align model and software releases so that new EXAONE generations can run on Furiosa RNGD as they become available, with adoption expanding on individual workloads are validated.
