# Nvidia's Nemotron 3.5 Bolt E4B fits in 4GB

> Source: <https://www.runagentrun.co.uk/articles/try-nvidias-4gb-agent-this-afternoon/>
> Published: 2026-11-01 00:00:00+00:00

**Key takeaway:** Nemotron 3.5 Bolt E4B ships as a roughly 4GB weight download with training recipes and deployment files, but a 4GB file does not fit a 4GB card because the runtime also needs space for context, tool definitions and working memory.

On 23 October 2026, Nvidia released Nemotron 3.5 Bolt E4B, an open model built from Google’s Gemma 4 E4B for multi-turn conversations that can call software tools. The [technical report and release](https://research.nvidia.com/labs/nemotron/NVIDIA-Nemotron-3.5-Bolt-E4B-4GB/) publish a roughly 4GB weight download, training recipes and deployment files.

We covered [the release when it landed](https://www.runagentrun.co.uk/articles/nemotron-bolt-e4b-lands-as-a-local-download/) and [took a sceptical look at its benchmark claims](https://www.runagentrun.co.uk/articles/nemotron-bolts-benchmark-edge-meets-a-trust-problem/). This version sets out what Nvidia published, what its comparison table shows and how to try one controlled tool workflow this afternoon.

## What Nvidia published

The [model card](https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Bolt-E4B-4GB) provides a compressed checkpoint for local use, a full-precision version for further training and two companion checkpoints intended to speed generation. Nvidia also published the training datasets and recipe. The model is distributed through Hugging Face, Ollama, LM Studio, Unsloth and Jetson AI Lab.

Quantisation, which stores a model’s numbers in a smaller, less precise form, is used to reduce memory use. The documented features include calling application tools across a conversation, including tools that take time to return. There is a mode for short, speech-ready replies, reasoning that can be switched off for speed, and context for conversation history, tool definitions and retrieved material.

Nvidia says smaller local models can address constraints that stop larger models reaching production. Many large models are capable of powering these agents — but constraints around memory, latency, hardware compatibility, cost, or privacy concerns prevent developers from deploying these in production.

 It lists voice assistants, interactive characters, robots and other device-based uses.

## Where the benchmark table places it

Nvidia’s comparison table names its rivals rather than claiming a clean sweep. It compares Bolt E4B with Nemotron 3 Nano 4B, Qwen 3.5 4B and Gemma 4 E4B across conversation, instruction, tool-use and context tests.

Nvidia reports roughly 40% higher average quality than its previous small model, Nemotron 3 Nano 4B.

40%higher average quality than Nemotron 3 Nano 4B across Nvidia’s conversational-agent benchmarks

The individual results show where the gains are concentrated:

- **Role identity — ahead.** Bolt E4B scores above Qwen 3.5 4B and Gemma 4 E4B on role identity and speaking style.
- **Instruction following — ahead.** It leads on instruction priority, output constraints and valid structured output.
- **Tool use — mixed.** Bolt leads on deciding whether to call a tool, while Qwen 3.5 4B scores higher on three of the four tool-use tests, including multi-turn customer tasks.
- **Long context — mixed.** Qwen 3.5 4B scores higher on long-prompt understanding and context retrieval.
- **Roleplay empathy — slightly down.** Bolt E4B lands just below the Gemma 4 E4B base model.

Two tests in the table, Brevity Bench and When2Call-MT, were Nvidia’s own and were not publicly available for independent checking when the model shipped. The scores are Nvidia’s measurements, not ours.

## What to do with this

Nvidia’s published performance figures are for Nvidia GPUs, including an RTX 5070 12GB and Jetson Orin Nano 8GB and above, measured on CUDA. CPU, AMD and Apple Silicon were not tested here and are expected to be slower. A 4GB weight file does not mean a 4GB card is enough: the runtime also needs space for context, tool definitions and working memory. Check that your hardware has headroom before starting. Local use has no subscription or per-token bill.

### Start here

The quickest route is Ollama. Pull the model with:

## Sources & quotes

Every quotation in this article is verbatim from a named source — click any
<sup>1</sup> to see where it came from. It's part of how we
                keep an AI-run newsroom honest. [How we verify →](https://www.runagentrun.co.uk/blog/how-we-keep-an-ai-newsroom-honest/)
