NVIDIA has released deployment recipes built on its Dynamo inference software for four of Korea's leading open models. The post, published on November 8, 2026 by NVIDIA's Amr Elmeleegy, covers A.X K2 from SK Telecom, K-EXAONE 2.0 from LG AI Research, Solar Open 2 from Upstage and Motif-3 from Motif Technologies.
The reasoning behind the release is a practical one. Once a model grows beyond what a single GPU can hold, it has to be split across several GPUs, and request routing and inter-GPU communication must be coordinated carefully to keep performance up. NVIDIA describes Dynamo as an open source, distributed inference-serving framework designed to handle that routing and communication for multi-node deployments at data center scale.
The recipes are optimized for NVIDIA B200 GPUs. NVIDIA says they remove much of the guesswork from deployment and deliver up to 2x performance out of the box, and that they should help developers bring Korea's open models into production sooner and turn them into broadly accessible services under Korea's AI for All initiative. The post is short and does not publish the benchmark setup, the baseline against which the 2x figure is measured, or per-model results, so the number should be read as an NVIDIA claim with an unspecified upper bound.
The release lands in a cluster of NVIDIA announcements around Korea. A companion post by Jesse Clayton on NVIDIA's expanded work with Korea's Ministry of Science and ICT says that, under AI for All, NVIDIA will work with three selected consortia and four leading Korean foundation model builders, using Dynamo to optimize and scale inference. The post does not list those four builders, and this package does not assume they are the same four models named in the recipes post.
A third NVIDIA post, by Nicola Sessions, adds a model-side detail: LG AI Research is opening API access for K-EXAONE-2.5, which it developed using NVIDIA Nemotron datasets, including Nemotron-Personas-Korea, an open dataset of 7 million fully synthetic Korean personas grounded in official Korean statistics. That is a newer version than the K-EXAONE 2.0 covered by the Dynamo recipes, so the two items describe different releases.
For teams running Korean-language or Korea-focused models, the practical value is a vendor-maintained starting point for multi-GPU serving rather than a new model or a new serving engine. The recipes target NVIDIA hardware only, and the post links to NVIDIA's general Dynamo materials for details.
Everything here comes from NVIDIA's own posts. No independent performance testing of the recipes, and no statements from SK Telecom, LG AI Research, Upstage or Motif Technologies about them, were reviewed for this package.