NVIDIA to Publish Dynamo-Optimized Recipes for Open Models From Four South Korean Organizations

NVIDIA announced Nov. 9, 2026, that it will publish deployment recipes optimized for NVIDIA Dynamo for open models developed by four South Korean organizations. The models are SK Telecom’s A.X K2, LG AI Research’s K-EXAONE 2.0, Upstage’s Solar Open 2 and Motif Technologies’ Motif-3.

NVIDIA says the four models require partitioning across multiple GPUs and adjustments to request routing when they exceed the capacity of a single GPU. Since releasing a model does not, by itself, determine the computing infrastructure configuration needed for production deployment, developers and operators will need to check the specific configurations in the Dynamo recipes and the scope of their application in real-world operations.

The focus of the announcement is the use of deployment recipes for NVIDIA Dynamo to work with four open models released by South Korean model developers. It is not a new announcement or performance update for the models themselves, but rather a move by NVIDIA to provide configuration and operational procedures for running them in inference environments. The inclusion of all four models suggests the announcement is intended to offer a framework for deploying different models, rather than a stand-alone installation for a specific model. However, publishing recipes does not demonstrate production use or a reduction in deployment time. The constraints cited by NVIDIA are important from a technical standpoint. Models that exceed the capacity of a single GPU must be partitioned across multiple GPUs, requiring adjustments that include request routing. As a result, releasing model weights alone does not complete preparations for real-world operations; the configuration of the inference infrastructure, including how processing is distributed among GPUs, is also an issue. The recipes are intended to configure this on Dynamo, but the specific settings and the number of GPUs required for each model cannot be confirmed from the text provided. The announcement should also be considered separately from reports about computing resources in South Korea, distinguishing supply from usage configuration. We reported on the outlook for delivery of contracted NVIDIA GPUs in South Korea expects no disruption to contracted NVIDIA GPU supply, vice science minister says. That report cited an explanation that 10,000 units funded through the 2025 supplementary budget were arriving and in operation, but there is no information that these recipes will be used with those GPUs or were designed based on the available supply. Securing GPUs and configuring models to run efficiently across multiple GPUs are separate stages of the challenge. The implication for the industry is that the connection between model developers and computing infrastructure providers may extend to deployment processes required after a model is released. It is not possible to determine how far recipes tailored for Dynamo standardize the usage requirements of each model, because no settings or measured results have been provided. Points to watch include the recommended GPU configuration for each model, performance and operating costs, the scope of the recipes and whether they are actually used in production. Without this information, the announcement establishes only which models are supported and what deployment challenges exist; it does not allow the impact of adoption to be assessed.

NVIDIA says models that exceed the capacity of a single GPU must be partitioned across multiple GPUs, and that request routing and other settings must be carefully adjusted. The recipes are optimized to deploy these four models on NVIDIA Dynamo. The announcement text provided does not specify the number of GPUs, performance figures or production deployment sites.