NVIDIA Nemotron Bolt E4B Guide: A Free 4GB AI That Talks on Almost Any GPU

Big AI usually lives in someone else’s data center and bills you per word. NVIDIA’s Nemotron 3.5 Bolt E4B 4GB, released October 23, 2026, is the opposite idea: a small, free AI designed to live inside a device — a robot, a game character, a shop kiosk — and talk without ever phoning home.

What it is, in plain language

Bolt E4B is an open-weight language model: 8 billion parameters in total, 4 billion “effective” (a distinction inherited from its base, Google’s Gemma 4 E4B — think of it as 8 billion parts, 4 billion doing the talking at any moment). NVIDIA post-trained it with reinforcement learning — the technique that rewards good answers instead of just copying examples — for fast, multi-turn conversation and tool use. It runs on 64K tokens of context (roughly 48,000 English words — plenty for a long conversation, not for a whole book).

The name tells the story: “Bolt” is NVIDIA’s family of efficient local models, and “E4B 4GB” is the promise — a conversational agent that fits in about 4GB of video memory. NVIDIA’s launch demo had it chatting and moving a Reachy Mini robot, running locally on an 8GB Jetson Orin Nano, with NVIDIA Cosmos handling vision. What this means for you: the words stay on your machine. No per-token bill, no data sent to a cloud, no monthly subscription — the model is free; the computer is the cost.

Price: $0 for the model

Item	Cost	What this means for you
Model weights	$0, OpenMDW-1.1 license, commercial use allowed	You can build a product on it and sell the product. The license is the reason this is a business tool, not just a toy.
Hardware	Whatever you already own (RTX card, Jetson, Spark)	If you own a gaming PC with a 4–5GB+ card, your marginal cost is electricity. If you don’t, this is not the project that justifies buying one.
Runtimes	$0 (Ollama, llama.cpp, LM Studio, vLLM)	All the software to run it is free and open.

Hands-on: running it in 15 minutes with Ollama

Step 1 — Install Ollama. Go to ollama.com, download it for your system, and open it. This is the free app that runs local AI models — think of it as a music player, but for models.

Step 2 — Pull the model. In a terminal, run:

ollama pull nemotron-3.5-bolt:e4b-q4_K_M

Ollama lists the e4b builds at about 7.2–7.9GB to download (the q4_K_M build plus its packaging). Wait for the download — this is the only slow step.

Step 3 — Talk to it. Run:

ollama run nemotron-3.5-bolt:e4b-q4_K_M

Step 4 — Try a job it is built for. Type:

You are the front-desk voice of a bike repair shop. A customer says: “My gears slip when I climb hills.” Ask one clarifying question, then give the two most likely causes in plain language.

Step 5 — What you should see. A fast, short, conversational answer — not an essay. That is the point: Bolt E4B is tuned for back-and-forth talk, where speed matters more than depth. If answers ramble, shorten your instruction; small models follow short instructions better than long ones.

Troubleshooting, honestly: the 4GB figure is a footprint goal, not the download size (6.18GB GGUF file, ~7.9GB via Ollama). NVIDIA lists Windows and Linux on Ada Lovelace and Blackwell GPUs; Mac support is not NVIDIA’s published path for this model, so Mac owners should expect extra work.

Make money with it: the $99/month local voice-agent setup

The customer: small shops that want a voice assistant (a kiosk, a phone greeter, a reception tablet) but will not send customer conversations to a cloud AI.

What you sell: a one-time setup ($299) plus $99/month care plan — Bolt E4B running on a mini-PC or Jetson they own, wired to their FAQ, with a simple script: greet, answer the top 20 questions, take a message when it can’t help.

Why it works: your model cost is $0 forever, so the $99/month is almost entirely margin for updates and support. The pitch that sells it is privacy: “your customers’ words never leave your shop.” Validate in 7 days: demo it on your own laptop at two local shops; one signed care plan proves the privacy angle pays.

Pitfalls

• Small means small. It beats its base model on most of NVIDIA’s internal tests (vendor claims, not independently verified) but trails Qwen3.5-4B on some tool-use tests. Don’t sell it as a replacement for a frontier model — sell it as a fast local talker.
• Text only. It does not see images or video on its own; the robot demo’s vision came from a separate model.
• This is a short guide, not a flagship review: Bolt E4B is a B-level release — free, useful, and narrow. If you need frontier quality, use a hosted flagship; if you need free, private, and fast, this is the pick.