# What Is Nemotron 3.5 Bolt E4B? Nvidia’s 4GB Local AI Model

URL: https://whatisagenticai.net/blog/what-is-nemotron-3-5-bolt-e4b/
Published: 2026-10-25
Publisher: Little Tough LLC, maker of AI Agent Assistant: Agentic AI (https://apps.apple.com/us/app/ai-agent-assistant-agentic-ai/id6761501260)

Nemotron 3.5 Bolt E4B is Nvidia’s tiny open model for local voice assistants and game characters, out Oct 23, 2026. What 4GB means, benchmarks, how to run it.

Nemotron 3.5 Bolt E4B is the smallest new model in Nvidia’s Nemotron 3.5 family, and its full name tells you its goal: **NVIDIA-Nemotron-3.5-Bolt-E4B-4GB**. Nvidia built it to run a talking assistant inside about 4GB of graphics memory, on a normal gaming PC or a small robot board.

If you saw "Nemotron Bolt E4B" or "Nemotron 3.5 Bolt 4GB" in a headline and wondered what it is, who it is for and whether you can use it, here is the plain-English version.

**TL;DR:** Nemotron 3.5 Bolt E4B is a free, open AI model from Nvidia, released on October 23, 2026. It is built on Google’s Gemma 4 E4B (8 billion total parameters, 4 billion effective) and retrained by Nvidia for voice assistants, game characters and robots that chat, call tools and stay in character. In Nvidia’s tests it used 3.1 to 5.0GB of graphics memory, depending on the device and how long the conversation was. You can download it from Hugging Face and run it in Ollama or LM Studio, but not on an iPhone.

## What is Nemotron 3.5 Bolt E4B?

It is a small language model, the text brain of an assistant. You type (or a speech tool turns your voice into text), and it writes back. Nvidia’s model card lists text in and text out only, with no image or audio input.

The name breaks down like this:

- **Nemotron 3.5** is Nvidia’s current family of open models for AI agents. Our [Nemotron 3.5 explainer](https://whatisagenticai.net/blog/what-is-nemotron-3-5/) covers the big members, Super VL and Bolt 74B-A9B, which came out on October 15.

- **Bolt** is Nvidia’s label for the versions squeezed to run locally.

- **E4B** comes from Google’s base model, Gemma 4 E4B. It means about 4 billion "effective" parameters out of 8 billion in total.

- **4GB** is the memory budget Nvidia designed it for.

The point is not to be the smartest model around. The point is to be quick and small enough to leave room on the same chip for speech recognition, a voice, a camera model or the game itself.

## Who made it, and when did it come out?

Nvidia made it. The model card and Nvidia’s research page both give the release date as October 23, 2026, on Hugging Face and build.nvidia.com. The Nemotron 3.5 family post we published on October 19 noted that Nvidia had promised this 4GB version "later in October." It arrived on schedule, and MIXED Reality News covered it on October 25.

Nvidia says the training data runs to May 2026, so the model does not know about anything after that.

## What does "4GB" really mean?

This is the part people get wrong. 4GB is a target, not a promise. Nvidia’s research page says the model fits into "as little as ~4GB in its target workloads." Here is what Nvidia measured with the 4-bit version:

| Device | Short chat (3K tokens in) | Long chat (32K tokens kept) | Speed, short chat |
| --- | --- | --- | --- |
| GeForce RTX 5070 (12GB) | 3.5GB | 4.3GB | 133 tokens/sec, first word in 0.55 sec |
| RTX Spark (32GB) | 3.1GB | 4.2GB | 67 tokens/sec |
| Jetson Orin Nano (8GB) | 4.2GB | 4.9GB | 15.5 tokens/sec, first word in 5.8 sec |
| Jetson AGX Orin (32GB) | 4.2GB | 5.0GB | 24 tokens/sec |

So a short chat on a gaming card stays under 4GB, and a long conversation goes a bit over. The download itself is bigger: the 4-bit GGUF file is 6.18GB. The 4GB figure is about graphics memory in use during Nvidia’s tests, not the file size.

## What is Nemotron 3.5 Bolt E4B good at?

Nvidia aims it at three jobs: voice and device assistants, interactive characters such as game NPCs, and agents inside robots and machines. It has a few switches for that work:

- **Conversational mode** makes replies short and easy to speak out loud.

- **Reasoning on or off.** Thinking is on by default. Turn it off for faster replies, or cap how long it thinks.

- **Tool calls**, so it can trigger an app action, like starting a robot’s wave animation.

- **A 64K-token context window** for long sessions.

- **A drafter model** called DFlash that Nvidia says gives up to 21% more speed than Gemma’s original drafter.

In one Nvidia demo, a Reachy Mini robot on a Jetson Orin Nano uses a vision model to see the room, and Bolt E4B picks which animation to play and writes what the robot says.

## Nemotron 3.5 Bolt E4B benchmarks

Nvidia ran the tests itself, and its card warns the numbers may differ from each vendor’s own scores. Here are a few from the 4-bit versions, which is what most people would run:

| Test (what it checks) | Bolt E4B | Gemma 4 E4B | Qwen3.5 4B | Nemotron 3 Nano 4B |
| --- | --- | --- | --- | --- |
| RoleBench (staying in character) | 71.2 | 60.5 | 52.4 | 10.9 |
| MultiChallenge (long multi-turn chats) | 55.6 | 39.4 | 52.2 | 34.5 |
| IHEval (following instruction priority) | 82.4 | 66.6 | 62.9 | 53.8 |
| BFCL-v3 (calling tools) | 67.3 | 65.0 | 70.6 | 59.1 |
| Tau2-Bench (agent tasks) | 52.3 | 38.7 | 61.7 | 14.9 |

The pattern is clear. Bolt E4B wins on chatting, role-play and following instructions, which is what a voice assistant or game character needs. Qwen3.5 4B still beats it on harder agent tasks and long-document tests. Nvidia says Bolt E4B averages about 40% higher than its older Nemotron 3 Nano 4B and uses up to 80% fewer output tokens when reasoning.

## Is Nemotron 3.5 Bolt E4B free?

Yes. The weights are free on Hugging Face under the OpenMDW-1.1 license, and the card says it is ready for commercial use. Because it is built on Gemma 4, the Apache 2.0 license on Gemma 4 also applies. Running it costs you only electricity and hardware.

## How to run Nemotron 3.5 Bolt E4B

- Search Hugging Face for **nvidia/NVIDIA-Nemotron-3.5-Bolt-E4B-4GB**. A full-precision BF16 version is also listed.

- Load the 4-bit GGUF file in a local app. The card lists Ollama, LM Studio, Jan, llama.cpp, vLLM, SGLang, Docker Model Runner and others.

- Use an Nvidia GPU for the speeds above. Nvidia tested GeForce RTX cards, RTX Spark, DGX Spark and Jetson boards.

It does not run on an iPhone. Nvidia does not offer an iOS version, and Apple phones do not use Nvidia chips.

## Nemotron 3.5 Bolt E4B vs Bolt 74B-A9B

|  | Bolt E4B 4GB | Bolt 74B-A9B 24GB |
| --- | --- | --- |
| Released | October 23, 2026 | October 15, 2026 |
| Size | 8B total, 4B effective | 74B total, 9B active |
| Memory target | About 4GB | 24GB |
| Built from | Google Gemma 4 E4B | Nemotron 3.5 Super VL |
| Best for | Voice assistants, game characters, robots | Local agents that browse, code and use a terminal |

Think of E4B as the fast talker and 74B-A9B as the worker. Local voice AI is moving fast in general: Block’s [Dessa voice model](https://whatisagenticai.net/blog/what-is-dessa-ai/) also aims to run on a laptop, but it is speech-to-speech and still a waitlist-only preview.

## What this means for you

Models like this are how talking characters in games, smart speakers and home robots will get smarter without sending every word to the cloud. That is good for speed and privacy. But it is a developer tool today. You need an Nvidia GPU, a download and some setup, and a 4B model will not research a topic or plan a trip as well as a big cloud model.

If you want an AI that finishes real tasks right now from your phone, an agent app is the shorter path. **Agentic AI** gives you a team of expert agents, like a Research Analyst and a Writing Coach, that work together on your iPhone and iPad. You describe the job, and they do the steps.

New to the words in this post? Our [agentic AI glossary](https://whatisagenticai.net/agentic-ai-glossary/) explains tool calling, context windows and more.

## Quick answers

**What is Nemotron 3.5 Bolt E4B?** A small, free AI model from Nvidia, released October 23, 2026, made for local voice assistants, game characters and robots. Its full name is NVIDIA-Nemotron-3.5-Bolt-E4B-4GB.

**Does Nemotron 3.5 Bolt E4B 4GB really fit in 4GB?** In Nvidia’s short-chat tests it used 3.1 to 4.2GB of graphics memory. Long chats with 32K tokens of history went up to 4.2 to 5.0GB.

**Is Nemotron 3.5 Bolt E4B based on Gemma?** Yes. Nvidia fine-tuned it from Google’s Gemma 4 E4B, which has 8 billion parameters and 4 billion effective.

**Can I run Nemotron 3.5 Bolt E4B in Ollama or LM Studio?** Yes, the model card lists both, plus llama.cpp, Jan and vLLM. A 4-bit GGUF file (6.18GB) is available.

**Can I use Nemotron 3.5 Bolt E4B on my iPhone?** No. It targets Nvidia GPUs and Jetson boards. For AI agents on iPhone, use an app like [Agentic AI](https://apps.apple.com/us/app/ai-agent-assistant-agentic-ai/id6761501260).

**Sources:** [Nvidia model card: Bolt E4B 4GB](https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Bolt-E4B-4GB) · [Nvidia research page](https://research.nvidia.com/labs/nemotron/NVIDIA-Nemotron-3.5-Bolt-E4B-4GB/) · [MIXED Reality News](https://mixed-news.com/en/nvidia-nemotron-bolt-local-ai-assistant-4gb/)