The development of artificial intelligence is rapidly surpassing the memory and software capacities of computers and workstations that developers rely on. Hence, they are often forced to conduct their research and prototyping in cloud environments or via large data centres.

With the development of the world’s smallest supercomputer, the NVIDIA DGX Spark, NVIDIA has resolved this issue. DGX Spark enables AI development directly in the laboratory or office.

The NVIDIA DGX Spark integrates the entire NVIDIA AI platform. The NVIDIA GB10 Grace Blackwell Superchip is connected to the CPU (central processing unit) via an NVIDIA NVLink C2C connection with 128 GB of coherent unified memory and the NVIDIA AI software stack, forming a powerful desktop AI supercomputer. This allows developers to build and test artificial intelligence models in a compact desktop form factor. The DGX Spark can locally run large language models with up to  200 billion parameters, transforming a desktop machine into a platform for the development of artificial intelligence.

Researchers Domen Vreš and Dr. Iztok Lebar Bajec from the Faculty of Computer and Information Science, University of Ljubljana, were awarded the DGX Spark through the NVIDIA Academic Grant Program. The program advances academic research by providing world-class computing access and resources to researchers. The project focuses on building an advanced English–Slovene translator based on the open-source GaMS-9B-Instruct model trained with NVIDIA NeMo. The model will be specialised via supervised fine-tuning (SFT), and DPO and GRPO will be evaluated using NeMo-RL to reduce factual hallucinations and improve instruction following. The optimised translator will be released as an open-source, production-ready NIM, along with a blueprint for other low-resourced languages.