At Domyn, we’re building multilingual models that can truly reason across math, code, tool use, and major European languages. And thanks to NVIDIA AI DGX Cloud Lepton, we’ve unified training, alignment, serving, and evaluation in one place.
With NeMo for CPT and SFT, NeMo RL for GRPO, and vLLM to spin high-quality synthetic data at scale, we can perform faster and more efficiently than ever. DGX Cloud Lepton lets us scale across many nodes, maintain high throughput and tighten our build–measure–learn loop, resulting in faster experiments, cleaner evaluations, and less operational drag. In short, it allows us to stay focused on what we do best without worrying about the underlying infrastructure.
Meanwhile, at the system level, NVIDIA DGX Cloud Lepton has been hands-on with performance tuning and training monitoring. The platform ties together vLLM for serving, and Dask and Ray for parallelism. Storage plugs into our S3 buckets and high-throughput file systems to keep pipelines moving.
Stay tuned for more exciting news from Domyn and NVIDIA!
Alexis Black Bjorlin, Yangqing Jia, Kourosh Behnam, Vishal Ganeriwala
