NVIDIA’s Nemotron Bolt E4B targets local agents on smaller hardware

NVIDIA’s model card dates the Nemotron 3.5 Bolt E4B 4GB release to October 23, 2026. The smaller local model is aimed at conversational agents, tool use and device assistants. This article was published October 27.

The details that matter

NVIDIA lists eight billion total parameters and four billion effective parameters, with the published GGUF using Q4_K_M quantization. The card describes a model post-trained from Gemma 4 E4B and lists a context of up to 64K tokens. Runtime, operating system, context size and accompanying tools all affect actual memory requirements; a 4 GB model designation is not a guarantee that every 4 GB device can run the complete workload.

NVIDIA documents several runtime paths on its model card, but also notes that its llama.cpp recipe is pending. Verify the current instructions and runtime compatibility before downloading. The model card lists the license as OpenMDW 1.1; it is not automatically the MIT license.

Our take: a small model for voice and device agents is interesting for the GPU-poor audience we serve, but we have not installed or tested this release and cannot promise speed, fit or launcher support.