NVIDIA released the previously announced Nemotron 3.5 Bolt E4B 4GB on October 23. Further trained from Gemma 4 E4B for multi-turn tool use and instruction following, it includes Q4_K_M GGUF weights and serving examples. NVIDIA describes eight billion total and four billion effective parameters. It can support local voice assistants, but the model itself takes and produces text, requiring a separate speech pipeline. The 4GB name also does not guarantee total memory use including long-context caches and runtime overhead, so deployment needs hardware-specific testing.