NVIDIA Releases Nemotron 3.5 Bolt E4B, a Local Agent Model for 4GB VRAM
The 8B-total/4B-effective SLM runs multi-turn tool-calling agents on RTX, Jetson and DGX Spark under the permissive OpenMDW-1.1 licence.
NVIDIA released Nemotron 3.5 Bolt E4B 4GB, a small language model with 8 billion total and 4 billion effective parameters, post-trained from Gemma 4 E4B using reinforcement learning only. The model targets voice assistants and multi-turn tool-calling agents running locally on RTX GPUs, Jetson devices and DGX Spark, with a 64K context window and reasoning on/off modes with a thinking budget.
The company reports the model outperforming Gemma 4 E4B and Qwen3.5-4B on 15 of 21 benchmarks, and running in about 3.5GB of VRAM on an RTX 5070 at Q4_K_M with 548ms time-to-first-token and 133 tokens per second. It is licensed under OpenMDW-1.1, which permits commercial use, with weights, training datasets and RL recipes released openly. Benchmark figures are vendor-reported and independent testers have noted score discrepancies between the model card and technical report.