NVIDIA describes it as a model for local voice and device assistants, interactive personas, and embedded interfaces, with improved multi-turn tool use, instruction following, and input grounding.
The release uses Q4_K_M precision and lists 8B total parameters, with 4B effective parameters. It supports up to 64K input tokens and 64K output tokens. The model card specifies text input and output, chat-template controls for thinking and conversational modes, and commercial as well as non-commercial use under the OpenMDW-1.1 license.
NVIDIA provides llama.cpp launch commands and a companion DFlash drafter for speculative decoding. A full-precision BF16 variant is also linked. The card says a llama.cpp recipe is pending.
The source includes NVIDIA-run benchmark comparisons against Gemma, Qwen, and an earlier Nemotron Nano model. It does not establish independent benchmark results, performance on a particular device, or production reliability for an agent workload.