NVIDIA added B200-optimized Dynamo deployment recipes for four Korean open models and said they can speed production deployment.

NVIDIA launches Dynamo recipes for four Korean models

NVIDIA launches Dynamo recipes for four Korean models

Image: AI-generated editorial illustration

NVIDIA said on Nov. 8, 2026, that it has added Dynamo-optimized deployment recipes for four Korean open models from SK Telecom, LG AI Research, Upstage and Motif Technologies. The company said the recipes are designed for B200 GPUs and are meant to simplify production deployment across multi-GPU systems.

What did NVIDIA announce on Nov. 8?

NVIDIA announced Dynamo-optimized deployment recipes for four Korean open models on Nov. 8, 2026. The company said the recipes cover 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 recipes are part of NVIDIA Dynamo, which NVIDIA describes as an open-source, distributed inference-serving framework built to deploy models across multiple nodes and GPUs at data-center scale. NVIDIA said the goal is to make it easier to bring Korean open models into production faster and to expand its AI for All services.

Why does NVIDIA say Dynamo matters for inference serving?

NVIDIA said larger models often exceed the memory or throughput capacity of a single GPU, which means they must be split across multiple GPUs. In that setup, the company said maintaining performance depends on coordinated request routing and communication.

According to NVIDIA, Dynamo handles that routing and communication for distributed inference and is built for multi-node deployment at data-center scale. In plain terms, the framework is meant to coordinate how requests move through a cluster so the model can serve traffic efficiently when one GPU is not enough.

NVIDIA repeated that message in its AI Day Seoul coverage, saying larger models require multi-GPU deployment and that Dynamo manages routing and communication for distributed inference.

Which Korean models got deployment recipes?

NVIDIA said the four models now have Dynamo recipes optimized for NVIDIA B200 GPUs: A.X K2, K-EXAONE 2.0, Solar Open 2 and Motif-3. The company said the recipes are intended to simplify production deployment of those models and help developers move them into production more quickly.

The documentation also shows how the recipes are organized in practice. NVIDIA’s Dynamo recipe catalog lets users select a model and a matching hardware or runtime target, then deploy the corresponding recipe. For version 1.6.0, the catalog listed 25 model families and 125 deployable configurations.

Among the supported configurations, NVIDIA lists an experimental Motif vLLM chat recipe for Motif-3 using two B200 GPUs, tensor parallelism, expert parallelism and Dynamo speculative decoding. NVIDIA also says the K-EXAONE 2.0 recipe targets a 750B-A37B mixture-of-experts model in NVFP4 on B200 hardware, with aggregated or disaggregated 1P1D serving and MTP speculative decoding on an 8K/1K chat trace.

For A.X-K2, the documentation says B200 profiles use NVFP4 weights and FP8 KV cache, while H200 profiles use FP8 weights and BF16 KV cache. NVIDIA says users can choose aggregated or disaggregated serving, with B200 deployment sizes of 8x or 12x B200 and H200 deployment size of 32x H200 for chat workloads.

What performance claims did NVIDIA make?

NVIDIA said the B200-optimized recipes can deliver up to two times performance out of the box, but the available materials do not provide an independent benchmark methodology or result. The company’s claim is therefore NVIDIA’s own performance assertion rather than a separately verified test result.

The documentation also indicates that the Motif-3 recipe is experimental, which suggests the catalog is still evolving even as NVIDIA pushes these configurations toward production use.

How does this fit into NVIDIA's broader Korea push?

NVIDIA said the recipes are part of a broader Korea collaboration tied to the AI for All initiative. The company said it will work with three selected consortia and four Korean foundation-model builders using Dynamo to optimize and scale inference.

NVIDIA and Korea’s Ministry of Science and ICT also said they are expanding collaboration across semiconductors, infrastructure, talent, startups, models, agents and security. The company said the collaboration includes training and certification programs, hackathons and technical seminars aimed at strengthening GPU utilization.

At the same time, NVIDIA said the Korea collaboration lays groundwork for Korean NPU and XPU companies to explore participation in the NVLink Fusion ecosystem. NVIDIA separately said it is collaborating with SK Telecom and Rebellions, with support from MSIT, on sovereign AI infrastructure using NVLink Fusion.

The Korea push has also been visible in NVIDIA’s Seoul event coverage. NVIDIA said AI Day Seoul brought together more than 2,000 developers, startups, enterprises and researchers at COEX, with sessions covering open models and agents for sovereign AI, physical AI, industrial transformation, AI for science, talent development, commercialization and national AI infrastructure. At the MSIT-NVIDIA AI Startup Partnership Forum, NVIDIA and Korea’s science ministry met with over 20 Korean startups, and NVIDIA said it invested in Point2 while planning additional investments in Motif Technologies, RLWRLD and Upstage.

Sources