The report describes it as further post-trained from Gemma4-E4B-it for multi-turn tool use, instruction following, input grounding, and structured output.
The release lists a Q4_K_M GGUF checkpoint, a BF16 Hugging Face-format checkpoint, and DFlash drafter variants for speculative decoding. NVIDIA also says the model is available through Hugging Face and NVIDIA NIM, with a free API endpoint and a local NIM download. The target memory footprint is about 4 GB.
In NVIDIA's RTX 5090 measurements, the Q4_K_M model recorded a 285 ms time to first token and a 70.2 representative benchmark average. With DFlash, NVIDIA reports 2.74x per-user decode speed versus non-speculative inference at concurrency one.
The report does not establish independent benchmark validation, performance on other hardware, production reliability, or the terms that govern every listed asset.