Tag
A full-stack evaluation of NVIDIA's Nemotron-3 Mamba-Transformer hybrid models on DGX Spark hardware, including architecture analysis, quantization (NVFP4), benchmark results, and introduction of the open-source smf-bench testing suite.
NVIDIA released the Nemotron 3 open model, offering three sizes: Nano, Super, and Ultra. It optimizes hardware efficiency through architectural innovations such as hybrid Mamba Transformer, latent MoE, and multi-token prediction, and adopts the Open MDW 1.1 open license.
NVIDIA releases Cosmos 3 (Mixture-of-Transformers models up to 64B), Nemotron 3 Ultra (550B-A55B LLM), and previews RTX Spark personal superchip at Computex 2026, achieving SOTA on multiple open model leaderboards.