NVIDIA says the 4 GB Nemotron Bolt model is available, and uses fewer tokens

Watch

Lab claim

On October 27, 2026, NVIDIA said Nemotron 3.5 Bolt E4B, the 4 GB model, is available for on-device game characters. NVIDIA says it is about 40 percent higher quality than its prior model on the benchmarks it chose, and that it writes up to about 80 percent fewer tokens than unnamed peers. The context window is 64,000 tokens.

At a glance

Claims are from NVIDIA's October 27 developer post. Comparison models are not named there.
Model	Nemotron 3.5 Bolt E4B
Size	4 GB
Context	64,000 tokens
Languages	35
Quality versus NVIDIA's prior model	About 40 percent higher, benchmarks not named
Tokens versus peers	Up to about 80 percent fewer, peers not named
Where it runs	Locally on GeForce RTX and RTX Spark, via NVIDIA's in-game SDK

How to read this claim

Measurement — The benchmarks behind the 40 percent quality claim are not named.
Bill — List input and output per million tokens on the vendor page — a loop pays both again.
Door — A public weight URL is not in the post. NVIDIA points developers to its in-game SDK.

Keep the habit · field card with prices and worked checks

What changed

The 4 GB Bolt model moved from an announced edge model to one NVIDIA says developers can use.

What it introduces

A 4 GB on-device Nemotron checkpoint with a 64,000 token context.

Who benefits Anyone building an on-device character or a local assistant that has to leave memory for the rest of the program.

Try this

If you need a small local model, the 4 GB Bolt is the one NVIDIA now says you can download through its in-game SDK. Do not treat the 80 percent token cut as a comparison until NVIDIA names the other model.

What we do not know

The benchmarks behind the 40 percent quality claim are not named.
The peers behind the 80 percent token claim are not named.
The post does not give a parameter count beyond the 4 GB label.
A public weight URL is not in the post. NVIDIA points developers to its in-game SDK.

Evidence

Primary Developer post. The 40 percent quality claim and the token claim are NVIDIA's. The comparison models are not named.

Independent The quality and token figures are NVIDIA's. The peers are not named, and no second lab has published a rerun.

Sources

Developer post Primary · October 27, 2026

The record

What happened

NVIDIA's October 27 developer post says Nemotron 3.5 Bolt E4B 4GB is available. It is aimed at in-game characters that run locally, with multi-turn tool use, a conversational mode, and reasoning that can be turned off. The context window is 64,000 tokens. NVIDIA says the model offers about 40 percent higher average quality than its prior model on relevant benchmarks, and outputs up to about 80 percent fewer tokens than peers. The post does not name those benchmarks or those peers. The same post describes RTX Spark, a Windows PC with up to 128 GB of unified memory, and an October DLSS 5 model update that NVIDIA says runs twice as fast as the September version, halves GPU overhead, and uses 20 percent less video memory at 4K.

Why it matters

A 4 GB local model is something a game can ship beside the graphics, if the quality claim holds. The claim is not checkable from the post, because the other models and the benchmark names are missing. What a reader can use today is the availability, the 4 GB size, and the 64,000 token context.

NVIDIA's October 15 post announced this 4 GB model and did not list it as available then. October 27 is the post that says it is available.

The DLSS 5 numbers, twice as fast and 20 percent less memory at 4K, are a graphics-model update in the same post. They are not scores for Bolt.