NVIDIA releases deployment recipes for four Korean AI models, optimized for B200 GPUs

NVIDIA is promising a faster path from Korean open AI models to working services. On November 8, 2026, it announced deployment recipes for four models using its Dynamo serving software. The recipes are optimized for B200 GPUs, with NVIDIA claiming up to twice the performance out of the box—a claim whose usefulness depends on the comparison.

Four models, one deployment release

The announcement concerns recipes for running the models, rather than the release of new models themselves. NVIDIA’s stated goal is to help developers bring these models into production faster. The package covers four named models from four Korean developers:

• A.X K2 from SK Telecom.
• K-EXAONE 2.0 from LG AI Research.
• Solar Open 2 from Upstage.
• Motif-3 from Motif Technologies.

Coordinating work across GPUs

The problem NVIDIA is addressing starts when a model grows beyond the capacity of a single GPU. Splitting it across several GPUs creates another job: coordinating where requests go and how the processors communicate. NVIDIA says that coordination requires care to maintain performance.

Dynamo handles that routing and communication. It is an open-source framework for inference serving—the process of running a model to produce responses. NVIDIA built it for deployments spread across multiple computing nodes at data-center scale. Its role here is to coordinate the serving system around the models.

What “up to 2x” leaves unresolved

NVIDIA does not identify the baseline behind its “up to 2x” figure in the announcement. It also does not specify the performance metric, test workload or results for each model. That leaves a material distinction unresolved: whether the gain means more output, shorter waits or another measure of serving performance. Developers have a hardware target and a model list, but not enough benchmark detail to translate the headline figure into an expected improvement for their own service.