Nvidia's Nemotron Bolt E4B: a tool-using agent in ~4GB of VRAM
Open weights on Gemma 4 E4B, GGUF on Ollama and LM Studio, 64K context, 35 languages — strong per gigabyte, but the agent benchmarks want a closer look.
Copy markdown
A 4GB model that runs real tool use
Nemotron 3.5 Bolt E4B is an open small model (built on Gemma 4 E4B) tuned for conversational agents — multi-turn tool calling, configurable reasoning, 64K context and 35 languages — at a measured 3.1–5.0GB peak VRAM. That fits a genuine agent loop on an 8GB laptop GPU or a Jetson.
Pull it tonight: GGUF, Ollama, LM Studio
The weights are up now: a post-trained GGUF plus a BF16 build on Hugging Face, with day-one support in Ollama, LM Studio, Unsloth and Jetson AI Lab. Weekend build: a private, fully offline tool-calling assistant that never phones home.
Read the fine print on the benchmarks
NVIDIA claims ~40% higher average scores than its own Nano 4B and up to 80% fewer reasoning tokens — but on agentic Tau2-Bench it trails Qwen 3.5 4B (55.1 vs 75.1) and on RULER-64K (87.0 vs 95.1). Reach for it for footprint and speed, not peak tool accuracy.
Free drafters for faster local tokens
The drop also bundles two DFlash drafter models for speculative decoding, so you can push more tokens/sec on the same card. Pair them with the GGUF for a snappier local loop.