Nemotron 3.5 Bolt E4B

NVIDIA Nemotron 3.5 Bolt E4B ('4GB') is the smallest Nemotron 3.5 Bolt model, announced 2026-10-15 with Super VL and Bolt 74B-A9B and released 2026-10-23. A post-trained Google Gemma-4-E4B-it: 8B total / 4B effective parameters, text-only, 65,536-token context, optional thinking mode, trained with asynchronous GRPO reinforcement learning for multi-turn conversation and tool calling. Built for local assistants, robots and game characters: ~3.1-5 GB VRAM depending on card and context (NVIDIA: 3.5 GB / 548 ms TTFT on an RTX 5070; runs on Jetson Orin Nano 8GB), with a companion DFlash speculative-decoding drafter. OpenMDW-1.1 weights (BF16, 4GB and NVFP4 variants) on Hugging Face, also via NVIDIA NIM; subject to Gemma terms for the base.