On October 23, 2026, NVIDIA released Nemotron 3.5 Bolt E4B in an official technical report — an open-source small language model billed as having the highest "intelligence density per GB." Built on Google's Gemma 4 E4B and refined with NVIDIA's Nemotron post-training, it ships with training datasets and the full recipe published openly.
An agent inside 4GB
Bolt E4B needs only about 4GB of VRAM under its target workload: the Q4_K_M quantized build uses 3.1–4.3GB — 3.5GB on an RTX 5070, 4.2GB on a Jetson Orin Nano 8GB (per official tests). It supports multi-turn tool calling (including guidance for asynchronous calls), a conversational_mode toggle (short, voice-friendly replies), configurable reasoning and thinking budget caps, 64K context, and 35 languages. NVIDIA claims its average quality on key conversational-agent benchmarks is about 40% higher than the previous Nano 4B, and that with reasoning enabled it emits up to 80% fewer output tokens than comparable models.
Against its peers: more wins than losses
Across the 21 comparison benchmarks NVIDIA published, Bolt E4B leads Gemma 4 E4B, Qwen 3.5 4B, and the previous Nano 4B in 15; it scores 84.8 on IHEval versus 67.0 and 60.6 for its rivals. It is the third member of the Nemotron 3.5 family — after the vision-oriented Super VL and the 74B-parameter Bolt 74B-A9B — and the smallest of the family.
Weaknesses and boundaries
Leading doesn't mean dominating. On Tau2-Bench, Bolt E4B scores 55.1, about 20 points below Qwen 3.5 4B's 75.1 — not the strongest on multi-turn conversational agent tasks. Every figure above comes from NVIDIA's official technical report, with no independent third-party verification yet; and 4GB is the quantized build's number, rising with higher precision or longer contexts.
Who should care
The model is already available through Hugging Face, Ollama, LM Studio, Unsloth, Canonical, Jetson AI Lab, and the NVIDIA In-Game Inferencing SDK. Demos include a Reachy Mini robot running on a Jetson Orin Nano 8GB — Cosmos 3 Edge handling vision, Bolt E4B handling voice and tool calls — and an in-game blacksmith NPC trade demo. With the recipe fully open, researchers and small teams can reproduce or fine-tune it directly — small models are moving from merely "runnable" to "built for agentic workflows."
