NVIDIA Dynamo Deployment Recipes Released for Developers
NVIDIA said it will release Dynamo-optimized deployment recipes for 4 Korean open models. GPU, storage, and runtime image requirements vary by model, and neither the comparison conditions for the claim of up to 2x performance nor operating costs have been confirmed.
- NVIDIA said it will release deployment recipes for A.X K2, K-EXAONE 2.0, Solar Open 2, and Motif-3.
- The B200 configuration for A.X-K2 requires 8 GPUs for integrated deployment and 12 GPUs for disaggregated deployment.
- K-EXAONE 2.0 requires a ReadWriteMany storage class with at least 700 GB of free space.
- Solar Open2 requires the dedicated image pinned in the recipe rather than the default Dynamo 1.4.1 vLLM runtime.
- The claim of up to 2x performance comes from NVIDIA; comparison conditions and independent verification results have not been confirmed.
NVIDIA announced that it is releasing Dynamo deployment recipes for Korean open models. The official page is dated November 8, 2026. The feed timestamp is November 9 at 14:00 Korean time. Developers can use the model-specific configurations for their own deployments.
This article follows NVIDIA's announcement and excerpts from official documentation and is current as of November 9, 2026.
Deployment Recipes for 4 Korean Open Models
NVIDIA's announcement covers 4 open models developed by Korean companies. The recipes provide configurations for deploying each model with Dynamo. NVIDIA said it optimized these recipes for B200 GPUs.
| Model | Developer | Configuration Confirmed in Official Documentation |
|---|---|---|
| A.X K2 | SK Telecom | Serving on B200 and H200 with Dynamo and vLLM |
| K-EXAONE 2.0 | LG AI Research | Serving the 750B-A37B model in NVFP4 format on B200 |
| Solar Open 2 | Upstage | Serving with a dedicated image pinned in the recipe |
| Motif-3 | Motif Technologies | An aggregated worker or a disaggregated configuration with KV-aware routing |
NVIDIA Dynamo is an open-source distributed inference serving framework. Inference serving means running a model so it can receive requests and return results. NVIDIA describes it as a tool for data-center-scale, multi-node deployments.
Model-Specific Requirements for Self-Deployment
Developers who plan to run a model themselves should check the resource requirements of their chosen recipe. The A.X-K2 documentation covers B200 and H200 configurations. The A.X-K2 GPU and memory figures below are requirements for the B200 configuration.
| Model | Configuration and Resource Requirements | Scope |
|---|---|---|
| A.X-K2 | 4 B200 GPUs and 400 GiB of host memory per worker | 8 GPUs total for aggregated deployment, 12 for disaggregated deployment |
| K-EXAONE 2.0 | A ReadWriteMany storage class with at least 700 GB of free space | Checkpoint disk usage is approximately 530 GB |
| Solar Open2 | 4 GPUs for 1 prefill worker, 4 GPUs total for 2 decode workers | 8 GPUs total for the disaggregated configuration |
| Motif-3 | 1 aggregated worker or a 3P1D configuration with 3 prefill workers and 1 decode worker | Each worker uses 2 GPUs |
Solar Open2's prefill configuration uses TP4 and expert parallelism. Each decode worker uses TP2. Motif-3 workers use TP2 and expert parallelism. Motif-3's MTP2 setting uses 2 speculative tokens.
These documents describe configurations for self-deployment. Usage conditions for Korea have not been confirmed. Whether hosting services are available in Korea is also not specified.
Recipe Fees and GPU Operating Costs
NVIDIA's announcement and the documentation excerpts do not list recipe fees. GPU acquisition costs and cloud operating costs are also not provided. Dynamo being open source does not mean that operating it is free.
- Recipe fees: Not specified.
- GPU and cloud operating costs: Not specified.
- Managed API availability: Not specified.
- Managed API request fees and limits: Not specified.
Up to 2x Performance and Requests Excluded from Evaluation
NVIDIA said the B200-optimized recipes deliver up to 2x performance with the default configuration. This is a performance claim made by the company. The official excerpts reviewed do not include the comparison baseline or measurement conditions. No independent verification results were provided.
| Item | What the Official Source States | Remaining Caveats |
|---|---|---|
| Recipe performance | NVIDIA's announcement claims up to 2x performance | Comparison baseline and measurement conditions unconfirmed |
| K-EXAONE 2.0 evaluation requests | 1,805 requests in NVIDIA's official evaluation documentation | 38 of these requests were excluded from the metrics |
| Reason for exclusion | HTTP 400 returned because requests exceeded the 262,144-token context limit | These failed requests are not reflected in the performance metrics |
The K-EXAONE 2.0 figures come from NVIDIA's official evaluation documentation. There is no basis for treating them as independent measurements. When citing the evaluation metrics, the request exclusion conditions should also be disclosed. The relationship between this evaluation and the up to 2x claim is also unconfirmed.
Exact Release Time and Service Scope Remain Unconfirmed
The difference between the official page date and the feed timestamp has not been explained. The official page displays November 8, 2026. The feed timestamp is November 9, 2026, at 05:00 UTC. Converted to Korean time, this is 14:00 on the same day.
This information alone cannot establish the exact release time. The time zone used for the official page date is also unconfirmed. The scope of managed services available in Korea has not been specified either.
- The reason for the date difference and the exact release time are unconfirmed.
- Usage conditions for Korea and the availability of hosting in Korea are unconfirmed.
- Managed API availability and request conditions are unconfirmed.
What to Check Before Deploying in Korea
For self-deployment, first check the runtime environment in each model's official documentation. Solar Open2, in particular, requires the specified dedicated image. The default Dynamo 1.4.1 vLLM runtime does not include this model's definition.
- A.X-K2 has different total GPU requirements for aggregated and disaggregated deployments. The B200 requirements should not be applied unchanged to the H200 configuration.
- A.X-K2 can reuse a PVC that already stores the model. A PVC is a request that specifies storage for a deployment. Set
claimNamein both the download job and the selected deployment source to that PVC. In this case, creating a new PVC can be skipped. - For K-EXAONE 2.0, the checkpoint size and required free space differ. Securing only approximately 530 GB does not meet the documented requirements. A ReadWriteMany storage class needs at least 700 GB of free space.
- Solar Open2 must use the image pinned in the recipe. It is not an optional component that can be replaced with the default runtime.
- For Motif-3, choose either an aggregated or a 3P1D deployment configuration. Check each worker's GPU requirements as well.
Before paying for cloud resources, check the cost estimate for your chosen environment. NVIDIA's up to 2x claim cannot be used as a performance guarantee for your own environment. Usage conditions for Korea and managed API support require separate confirmation.
FAQ
Has a managed API that can be called directly from Korea been released?
The confirmed announcement covers a Dynamo recipe for deploying the model yourself. The availability of a managed API, call pricing, and usage limits were not specified. Terms of use in Korea and the availability of hosting in Korea have not been confirmed either.
How many GPUs are needed to deploy A.X K2?
The B200 configurations in the official documentation require 8 GPUs for the aggregated setup and 12 GPUs for the disaggregated setup. Each worker requires 4 B200 GPUs and 400 GiB of host memory. H200 configurations are also documented, but there is no basis for applying these figures as H200 requirements.
Can I expect up to 2 times the performance and free use even if I run it in Korea?
NVIDIA claims up to 2 times the performance. The comparison baseline, measurement conditions, and independent verification results have not been confirmed. Recipe usage fees and GPU and cloud operating costs were not specified either.
Sources
Images
