Nvidia released Dynamo deployment recipes for four South Korean open models on 8 November, saying the B200-optimised software delivers up to 2x performance for inference across multiple GPUs, according to a company blog post announcing the release.
Key points
- The recipes cover models from SK Telecom, LG AI Research, Upstage and Motif Technologies.
- Nvidia says Dynamo coordinates requests and communication when inference runs across multiple GPUs.
- LG AI Research is reviewing Dynamo for its own model operations.
- LG is separately making a newer K-EXAONE model available through a free API.
Dynamo recipes run on B200 GPUs
The release covers 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. Nvidia describes Dynamo as an open-source framework for serving AI models across multiple machines in a data centre. The recipes are optimised for its B200 GPUs.
A model too large for one GPU has to run across several, Nvidia says. That requires the serving system to direct incoming requests and coordinate communication between GPUs. Dynamo handles those tasks across multiple machines, according to the company. The recipes provide deployment configurations for the four named models rather than new versions of the models themselves.
Nvidia says the configurations deliver up to 2x performance out of the box and can help developers move the models into production faster. The claim concerns inference serving: running a trained model to generate responses. It is separate from claims about a model’s ability to complete a task. Nvidia positions the recipes for deployments in which a single GPU cannot hold the model.
LG AI Research reviews Dynamo use
LG AI Research said it was reviewing whether to apply Dynamo to improve its model’s inference speed and operating efficiency, Digital Today reported on 9 November. Its comments accompanied an announcement at Nvidia AI Day Seoul 2026, held at COEX in Seoul, about access to a newer model, K-EXAONE 2.5.
The distinction matters for the two announcements. Nvidia’s deployment recipe names K-EXAONE 2.0. LG describes K-EXAONE 2.5 as a technology-verification model developed through additional post-training of K-EXAONE 2.0. LG said it would make the newer model’s API free to use from 9 November through its EXAONE API platform.
Developers, startups, companies and research institutions can obtain an API key for service development and research without additional cost, according to LG’s announcement. The institute plans to use that access to assess the model in operating services and broaden its use in South Korea. Its review of Dynamo concerns inference operations, while the free API concerns access to K-EXAONE 2.5.
K-EXAONE 2.5 enters free API access
LG AI Research says K-EXAONE 2.5 performed more than 70 percent better than the model released in August on a combined assessment of 24 evaluation metrics. That is LG’s assessment of model performance, rather than Nvidia’s claim about the speed of the Dynamo deployment recipes.
In LG’s internal evaluation using public datasets, K-EXAONE 2.5 scored 74.6 on AutomationBench, which tests work automation using software services. LG put DeepSeek V4.1 Flash at 74.9 and GLM 5.3 Max at 71.6 on the same measure. The company also reported scores of 78.0 on BrowseComp, a web-browsing test, and 59.2 on SWE-Bench Pro, a software-development test.
LG AI Research plans to apply post-training techniques tested with K-EXAONE 2.5 to the next phase of its foundation-model project. LG AI Research aims to complete an EXAONE model with what it calls frontier-class performance by January 2027.
