NVIDIA has released NVIDIA Dynamo recipes for four of Korea’s leading open models, giving developers a ready-made path to run very large open models in production. 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, and NVIDIA says they deliver up to twice the performance out of the box on its B200 GPUs.
The four models in this launch range from 250 billion to 750 billion total parameters. As models grow beyond the capacity of a single GPU, they must be split across multiple GPUs, and keeping that spread fast is the job NVIDIA Dynamo was built for. The recipes apply Dynamo’s serving machinery to these four models, tuned for NVIDIA’s B200 GPUs.
What NVIDIA released
NVIDIA announced the recipes on 9 November 2026. Each recipe is a Dynamo-optimized deployment configuration for one of the four Korean open models, tuned for NVIDIA’s B200 GPUs. NVIDIA’s claim for them is simple: they “take the guesswork out of deployment, delivering up to 2x performance out of the box.”
The stated goal is speed to production. The recipes are meant to help developers bring Korea’s open models into production faster and, in NVIDIA’s words, “turn them into broadly accessible AI for All services.” The announcement itself is short on detail about what a recipe physically contains, and that is one of the open questions this article returns to below.
What NVIDIA Dynamo is
NVIDIA Dynamo is an open-source, distributed inference-serving framework built to deploy models in multi-node environments at data center scale. It is not itself an inference engine. It sits above engines such as SGLang, TensorRT LLM and vLLM, and coordinates their workers across many GPUs and many machines.
Dynamo’s core idea is disaggregated serving: it splits the two phases of language-model inference — prefill, where the model processes the prompt, and decode, where it generates the response token by token — across different GPUs, so each phase can be scaled and optimized independently. It also routes each incoming request to the GPU that already holds the most relevant KV cache, the stored attention state from earlier tokens, so work is not recomputed needlessly.
Two components do most of this work. The KV-aware Router directs traffic across large GPU fleets to minimize redundant KV cache re-computation. The SLO Planner monitors capacity and prefill activity across nodes and adjusts GPU resources to keep latency within the service level objectives a team has set. Dynamo also extends GPU memory by moving KV cache data to cheaper storage tiers when it does not fit on the GPUs.
NVIDIA describes Dynamo as the successor to the Triton Inference Server, its earlier open-source serving software, and says Dynamo scales from a single GPU to thousands of GPUs. NVIDIA NIM microservices will include Dynamo capabilities, and Dynamo will also be supported in NVIDIA AI Enterprise, giving teams a packaged route to the same serving stack.
Why large open models need multi-GPU serving
As models grow beyond the capacity of a single GPU, they must be split across multiple GPUs, and the split model has to be served as one system. That is the problem Dynamo addresses: the framework handles the routing and communication between GPUs so inference performance holds up at data center scale.
That coordination is where performance is won or lost. As NVIDIA puts it, splitting a model across GPUs requires “careful coordination of request routing and communication to maintain performance.” The routing has to send each request to the GPU best placed to serve it, and the prompt-processing and token-generation phases have to be kept from competing for the same hardware — the two jobs Dynamo separates across different GPUs so each can be optimized on its own.
Done badly, a multi-GPU deployment spends its time waiting on communication and recomputing state it already held. Done well, the GPUs behave like one system, and that is what a serving framework is for: the routing, the phase separation and the cache management are the product, not an afterthought to the model itself.
What the NVIDIA Dynamo recipes deliver
The recipes apply Dynamo’s machinery to four specific models, optimized for B200 GPUs. What NVIDIA has said is that they make deployment faster and take the guesswork out of it; what it has not said is what a recipe contains, so a team picking one up cannot yet tell from the announcement which deployment decisions it settles and which remain theirs.
The headline number is “up to 2x performance out of the box,” attributed to NVIDIA. The announcement gives no workload, no baseline and no measurement method for that figure, so it should be read as NVIDIA’s claim for its optimized recipes, not as a guarantee for any particular workload — and teams should measure the gain on their own traffic before counting on it.
The models themselves are open and freely available. The serving challenge — routing requests, separating inference phases, managing the KV cache across many GPUs — is what Dynamo exists to handle, and the recipes apply that machinery to these four models.
The four models and their makers
The four models are all large, openly licensed models released in the summer of 2026, and three of them are finalists in South Korea’s government-backed Sovereign AI Foundation Model Project, which is building domestic foundation models. The sizes and dates below come from a tracker of Korea’s frontier models, based on the makers’ own statements.
A.X K2, from SK Telecom. At 688 billion total parameters, with 33 billion active, released on 29 July 2026 under the Apache 2.0 licence. SK Telecom says it trained the model from scratch on 8.2 trillion tokens, and that it improves on its predecessor, A.X K1, by an average of 32.2 percentage points across 14 benchmarks. It is available on Hugging Face.
K-EXAONE 2.0, from LG AI Research. At 750 billion total parameters, with 37 billion active, it is the largest model built in Korea, according to the tracker. Released on 31 July 2026 under the Apache 2.0 licence.
Solar Open 2, from Upstage. The smallest of the four at 250 billion total parameters, with 15 billion active, released on 22 July 2026 under the Upstage Solar License, an Apache-based licence.
Motif-3, from Motif Technologies. At 314 billion total parameters, with 13.2 billion active, released in August 2026 under the MIT licence. Motif says it built the model from scratch on about 12.5 trillion tokens. Its model is open and included in NVIDIA’s release.
Together the four span Korea’s open-model makers: SK Telecom, LG AI Research and Upstage — the three finalists in the Sovereign AI Foundation Model Project — plus Motif Technologies. Whether NVIDIA chose these four to cover the country’s open-model effort, or for another reason, is not something the announcement says.
What this means for production deployment
For teams running open models, the recipes aim to shorten the distance between downloading weights and serving users. NVIDIA says they take the guesswork out of deployment and bring the models into production faster; what the announcement does not say is what a recipe contains, so how much of the deployment work it settles for a team is still an open question.
The effect compounds for the models’ makers. NVIDIA frames the recipes as a way to bring Korea’s open models into production faster and turn them into broadly accessible “AI for All” services — open weights that a Korean company, agency or developer can run for its users.
Dynamo itself is open source and works with several inference engines, while the recipes are tuned for NVIDIA’s B200 GPUs. Dynamo’s capabilities are being folded into NVIDIA NIM microservices and NVIDIA AI Enterprise, giving teams a packaged route to the same serving stack.
Limits and open questions
The announcement leaves several questions unanswered. NVIDIA has not described what a recipe contains — whether it is configuration files, container images, scripts, or all three — or where exactly developers fetch it. It has not said whether recipes will be available for GPUs other than the B200, which matters for teams running older or different hardware.
The “up to 2x” performance claim is also unanchored: no workload, no baseline and no measurement method were published with it. Until those details appear, the claim should be read as NVIDIA’s own figure for its optimized recipes, and teams should measure the gain on their own traffic before counting on it.
Finally, it is not clear whether the four Korean models are the start of a broader recipe programme or a one-off tied to Korea’s Sovereign AI push. Would NVIDIA extend recipes to other open models, and would pre-tuned serving for a country’s flagship open weights become a standard route to production? The announcement does not say — that is the open question the launch leaves behind.
