Verified, then logged.
Sun 22 Nov 2026 · 00:34 KST / 15:34 UTC한국어 | ENSearch /

08 / Industry · Analysis

Nvidia publishes Dynamo recipes for four South Korean models

GPUs and configurations differ by model in Nvidia's documents, and the scope of the reported performance gain of up to twice is not yet disclosed

Four adjacent GPU server racks, each cabled differently, stand in a data center while a technician seen from behind works at a laptop in the background. AI-generated image

Nvidia said on 9 Nov 2026 that it is releasing recipes for running four Korean open models on Dynamo, its distributed inference framework, according to AI Times, a Korean AI news outlet. The reported performance gain is up to twice on B200 GPUs. The models come from the mobile carrier SK Telecom, LG AI Research, the AI arm of the LG group, and the Korean AI developers Upstage and Motif Technologies.

A recipe is a deployment procedure that sets out the configuration needed to run a model. AI Times presented the reduced burden of complex system configuration as the core of the release. The performance figure and the actual running conditions still need to be read separately.

Nvidia's deployment documents list a separate running environment for each model. How far the reported maximum of twice applies is also an open point. That makes the differences in GPUs and configurations across the documents the place to start.

Running conditions differ from model to model

Nvidia's Korean-language product page describes Dynamo as an open-source distributed inference serving framework. Inference is the work of running a trained model to produce answers. The page gives Dynamo's intended scope as multi-node environments at data-centre scale.

The A.X-K2 document specifies serving "on B200 or H200". It covers both aggregated and disaggregated setups with vLLM, an open-source model-serving engine; in a disaggregated setup, the stages of handling a request run on separate workers. The K-EXAONE 2.0 document states that the 750B-A37B model in NVFP4 format is served with vLLM on B200.

The Solar Open2 document covers serving Solar-Open2-250B-NVFP4 on B200 in either aggregated or disaggregated form. The Motif-3 document gives its configuration as "aggregated or 3P1D KV-aware" and names B200 as the GPU. Of the four documents, only A.X-K2 also names H200 in its description.

What the figure of up to twice covers

AI Times reported the gain of up to twice in inference performance as a figure for B200 GPU environments. B200 appears in the running environment of all four deployment documents. H200 appears only in the A.X-K2 document, so the scope of the figure needs separate scrutiny.

The AI Times wording points to a maximum and gives no basis for reading it as an average across configurations. Public documents do not show the existing setup used as the baseline, or whether the figure applies to each model. The number is therefore best read as a reported ceiling on B200.

Operator decisions that remain after the release

The deployment procedures for the models AI Times described as developed by Korean companies now appear model by model in Nvidia's documentation. The documents sit under a dev version path. Their descriptions give no schedule for moving them to the official documentation.

AI Times summed up the effect of the release as lighter configuration work and higher performance. Public documents do not show how large a gain an operator would get from a chosen configuration. Choosing the GPU and configuration remains the operator's job.