NVIDIA Makes Dynamo Recipes Available for Four Korean Open Models

NVIDIA Dynamo recipes are now available for four Korean open models: SK Telecom A.X-K2, LG AI Research K-EXAONE 2.0, Upstage Solar Open2 250B and Motif Technologies Motif-3. NVIDIA announced them on November 9, 2026, and made them available in the Dynamo platform documentation and the relevant repository.

This is not the release of new language models. The published materials are intended to help teams operate them on multi-GPU infrastructure using NVIDIA Dynamo and the vLLM runtime. The catalog lists supported runtimes, hardware configurations and deployment topologies.

NVIDIA Dynamo recipes for four models

Separate recipes are available for A.X-K2, K-EXAONE 2.0, Solar Open2 250B and Motif-3. For the first three models, the repository labels the configurations as complete deployments and also includes benchmark manifests. These are intended to provide a specific foundation for repeatable testing and configuration rollout.

Motif-3 is listed differently in the documentation. vLLM recipes are available for two or eight NVIDIA B200 GPUs, but their status is experimental. This distinguishes it from the other three models, whose recipes are labeled complete deployments.

The documentation also uses configurations for B200 and H200 hardware, depending on the specific recipe. The publicly available materials therefore do not represent a general guide that can be used without modifications on any setup, but predefined configurations for the listed combinations of models, runtimes and infrastructure.

Why the configurations matter

Operating large models, including mixture-of-experts (MoE) architectures, involves coordinating multiple GPUs, working with KV cache and routing incoming requests. Deployment recipes bring together the settings needed for these operational components into practical artifacts that can serve as a starting point for custom implementation.

For teams deploying these models, the published configurations could reduce the work involved in preparing an NVIDIA Dynamo and vLLM environment. The benchmark manifests for A.X-K2, K-EXAONE 2.0 and Solar Open2 250B also provide a structure for testing these deployments.

However, the availability of the recipes alone does not indicate what performance or cost benefit a particular organization will achieve. Results depend on the hardware used, the operating profile and the configuration selected. The available materials also do not independently establish that all of these configurations are already being used by customers in real-world production.

Motif-3 remains experimental for now

The experimental designation of the Motif-3 recipes is important to consider. They are available for vLLM on two or eight B200 GPUs, but the documentation does not place them among configurations at the same level of maturity as the recipes for A.X-K2, K-EXAONE 2.0 and Solar Open2 250B.

Another signal will be whether NVIDIA or the individual model creators publish independently reproduced benchmarks or information about customer deployments. It will also be interesting to see whether Motif-3 is eventually upgraded from experimental status to a fully validated configuration and whether support expands to additional hardware or runtimes beyond the currently listed B200 and H200 options.

Verified and updated: 11/09/2026 06:23