NVIDIA Dynamo Recipes Bring Korea’s Leading Open Models to Production

NVIDIA is shipping configuration rather than new software: Kubernetes deployment recipes for serving four large Korean open-weight models on Dynamo, its open-source framework for spreading one model across many GPUs. The models are SK Telecom's A.X K2, LG AI Research's K-EXAONE 2.0, Upstage's Solar Open 2 and Motif Technologies' Motif-3, all of which took part in South Korea's government contest to build a national frontier model (earlier coverage). Each recipe, listed in the ai-dynamo/dynamo repository's recipe catalog, fixes the GPU count, how the model is split across those GPUs, the number formats and the serving options, so a team does not have to work them out by trial and error, and each comes with a benchmark job for checking the result on your own cluster. The catalog introduces its recipes as "production-tested", but it lists Solar Open 2 and Motif-3 under "Experimental Recipes", which it describes as functional but "not yet fully validated for production use". Only the A.X K2 and K-EXAONE 2.0 recipes sit in its main table.

K-EXAONE 2.0 shows what those choices involve. It is a 764.5-billion-parameter mixture-of-experts model that uses about 37 billion parameters for each token. The recipe stores its weights in NVFP4, a 4-bit number format, keeps the attention cache (the stored intermediate results that spare the model from re-reading the conversation) in 8-bit FP8, and runs the whole model on four NVIDIA B200 GPUs. A second, disaggregated version uses eight B200s: four GPUs read the incoming prompt and build the cache, four others generate the reply, and the cache moves between them over InfiniBand. The other recipes follow the same pattern on B200s: A.X K2 on 8 or 12, Solar Open 2 on 4 or 8, and Motif-3 on 2 or 8. A.X K2 also has two versions for 32 H200 GPUs. A.X K2, Motif-3's eight-GPU version and Solar Open 2's four-GPU version use KV-aware routing, which sends each request to the worker that already holds cached work for the same opening text instead of rotating requests among workers in turn. Solar Open 2's eight-GPU version rotates requests on purpose, because it has a single prompt-reading worker and so nothing for the router to choose between, and K-EXAONE 2.0 ships KV-aware routing switched on but idle until a team runs more than one copy of the model.

The blog post's headline number, "up to 2x performance out of the box", has no stated baseline, and the repository does not supply one. The catalog keeps a separate set of "feature comparison" recipes that ship a baseline beside the optimized setup, and none of the four Korean models is among them. Each model's recipe documentation does publish measured throughput, but only for that recipe's own versions on NVIDIA's test traffic, never for a recipe against a hand-built deployment. For K-EXAONE 2.0 the page reports 87 output tokens per second per GPU on four B200s and 85 on eight, and says the eight-GPU version is for lower per-token latency and separate scaling of the two GPU groups, not for throughput. The A.X K2 benchmark guide explains how to compare KV-aware routing fairly against simple rotation, starting both runs from empty caches, but no result of that comparison is published. The published numbers also carry conditions. K-EXAONE 2.0's test traffic was designed for about 70% prompt reuse but achieved about 8.8% on the four-GPU version, because the cache could hold only a small part of the working set. The Motif-3 figures and A.X K2's B200 figures were measured with speculative decoding (drafting several tokens ahead, then checking them) forced to a fixed acceptance rate, which the Motif-3 page says does not measure real acceptance or output quality.

For a team that already runs Dynamo on Kubernetes with B200 GPUs, the recipes are a reproducible starting configuration for models that need several GPUs, with two of the four still marked experimental. The catalog assumes the Dynamo Kubernetes platform is installed first, a layer NVIDIA covers separately (earlier coverage). The 2x figure should be treated as NVIDIA's unverified claim until it says what the recipes were measured against.