Holotron-12B - High Throughput Computer Use Agent

Holotron-12B is a new multimodal computer-use model post-trained from the NVIDIA Nemotron-Nano-2 VL model. It is designed to serve as a policy model for computer agents that perceive, decide, and act in interactive environments. The model utilizes a hybrid State-Space Model (SSM) and attention architecture, which is optimized for high-throughput inference and handling long contexts with multiple images. When evaluated on agentic workloads, Holotron-12B achieves over double the throughput of a comparable transformer-based model on the same hardware, making it efficient for production use.