Bolt E4B is Gemma 4 E4B with NVIDIA's post-training for multi-turn tool use, instruction following and short, speech-ready replies. You can switch reasoning on for hard requests or off for speed, and there's a conversational mode that keeps answers brief. It covers 35 languages.
NVIDIA's numbers (self-reported, Q4_K_M on llama.cpp): 3.5GB of VRAM and 548 ms to first token on an RTX 5070 for a 3K-token prompt, about 133 tokens a second. On a Jetson Orin Nano 8GB it's much slower, near 5.8 seconds to first token. NVIDIA says it scores about 40% higher than its older Nemotron 3 Nano 4B across its chosen benchmarks and writes up to 80% fewer tokens than peers with reasoning on. It isn't ahead everywhere: Qwen 3.5 4B beats it on Tau2-Bench, 75.1 to 55.1.
Ollama v0.40.10 ships it with a DFlash draft model for speculative decoding. One thing doesn't line up: NVIDIA's page says 64K context, while Ollama's release notes say 128K.