Futureweb

Artificial Intelligence

NVIDIA's research page dated October 23 publishes Nemotron 3.5 Bolt E4B, a model aimed at about 4 gigabytes. An October 15 post said it was coming that day. The quality claims are NVIDIA's.

NVIDIA's October 23 page publishes Nemotron 3.5 Bolt E4B, aimed at about 4GB. An October 15 post scheduled that day. The quality claims are NVIDIA's.

The short version

A small model, built on Gemma 4 E4B with Nemotron post-training, aimed at about 4 gigabytes, with a 64,000-token context. NVIDIA says it targets GeForce RTX, Jetson, RTX Spark, and DGX Spark. The page's quality and token-reduction figures are NVIDIA's comparisons. An October 15 post said this variant was coming on October 23. MarkTechPost is an account of the page, not a second lab.

October 23 is the page

NVIDIA's research page for Nemotron 3.5 Bolt E4B 4GB is dated October 23, 2026. It says the model builds on Gemma 4 E4B with Nemotron post-training, and that the target is about 4 gigabytes. It names GeForce RTX, Jetson, RTX Spark, and DGX Spark as the machines it has in mind. Context is 64,000 tokens. The page says the model covers 35 languages, and that the base was pretrained on more than 140. It describes multi-turn tool use, a conversational mode, and configurable reasoning. Those are the page's features. MarkTechPost's piece is an account of this release. It is not a second measurement.Dessa's preview does not print a size.

October 15 scheduled it

NVIDIA's October 15 blog about Nemotron 3.5 Super VL said Bolt E4B was coming on October 23, and it treated other variants, including a larger Bolt, as already available that day. The October 15 post is a calendar. It is not the E4B release. A reader who files the whole family as shipped on October 15 will be early for this one. A reader who files the research page as a surprise will miss the appointment the blog already printed.Phonon publishes a voice size of a different kind.

The percentages are the company's

The research page says average quality is about 40 percent higher than NVIDIA's prior model on the conversational-agent benchmarks it is using, and that the model uses up to about 80 percent fewer tokens than peers. Both sentences belong to NVIDIA. This page did not run the suite and does not name a peer the page does not name. Up to is doing work in the token sentence. A single prompt can miss it. A 4 gigabyte target can also miss a board with less free memory than the slogan assumes. The page is a specification sheet from the maker, which is what it should be, and not an audit.Kimi's post is a different company's numbers.

What 4 gigabytes decides

A size target decides where the model can sit: a PC GPU, a Jetson, a small spark box, rather than a rack the reader does not own. It does not decide that the answers are safe, current, or cheaper than a hosted model once tokens and mistakes are counted. NVIDIA's token claim is the company's argument on cost. It is an argument about peers on the company's charts. Local inference keeps the weights on the machine, which is a privacy property only for the prompts that never leave. The page is not a policy.on-device inference keeps a specific set of things.

Languages, without collapsing lists

Thirty-five languages, and a base pretrained on more than 140, are the research page's figures. A container card or a catalog that lists a shorter set of languages is a different document, and this brief does not silently average them. If a download page names six languages, that page is the card for that download. The research page remains the one dated October 23. Readers should not assume the lists match.

Who is in scope

People who can run a model of about this size on the hardware NVIDIA names, and who want the company's description of a conversational agent on device. People waiting on a hosted API are not the audience of a 4 gigabyte target. MarkTechPost's readers are the audience of an account. The account does not replace the page. Asked directly, Did the October 15 blog ship Bolt E4B? It said the variant was coming on October 23, alongside other Nemotron models that were already available. The research page dated October 23 is the release this brief uses.

What this page will not claim

It will not claim the October 15 blog was the ship date. It will not treat 40 percent or 80 percent as this newsroom's measurement. It will not call a small local model a safety proof. It will not mix this variant with the larger Bolt the October 15 post said was already out. Same family, different file, different day. Asked directly, Whose number is the 40 percent? NVIDIA's. The page describes about 40 percent higher average quality than its prior model on conversational-agent benchmarks the company is using. This page did not rerun them.Canva's planning feature is a different ship.

Gemma underneath

Building on Gemma 4 E4B means the starting weights are that model, and Nemotron post-training is what NVIDIA says it added. A reader still needs Gemma's license for the base and NVIDIA's terms for the result. This page does not restate either license. It notes that a post-trained model has two parents, and that the October 23 page is the child. Asked directly, And the token claim? NVIDIA says up to about 80 percent fewer tokens than peers. Up to is the company's hedge. It is not a guarantee on a reader's prompt.

The limit

The record is a research page dated October 23, a blog that scheduled that date, and an outside account. The numbers on the page stay NVIDIA's. The short answer this page will stand on is this: A small model, built on Gemma 4 E4B with Nemotron post-training, aimed at about 4 gigabytes, with a 64,000-token context. NVIDIA says it targets GeForce RTX, Jetson, RTX Spark, and DGX Spark. The page's quality and token-reduction figures are NVIDIA's comparisons. An October 15 post said this variant was coming on October 23. MarkTechPost is an account of the page, not a second lab.

Sources

The reports this brief is filing. Futureweb did not republish them.

  1. NVIDIA Research
  2. NVIDIA blog, October 15
  3. MarkTechPost

Questions

Did the October 15 blog ship Bolt E4B?

It said the variant was coming on October 23, alongside other Nemotron models that were already available. The research page dated October 23 is the release this brief uses.

Whose number is the 40 percent?

NVIDIA's. The page describes about 40 percent higher average quality than its prior model on conversational-agent benchmarks the company is using. This page did not rerun them.

And the token claim?

NVIDIA says up to about 80 percent fewer tokens than peers. Up to is the company's hedge. It is not a guarantee on a reader's prompt.

What hardware does the page name?

GeForce RTX, Jetson, RTX Spark, and DGX Spark, for a model aimed at about 4 gigabytes. A target size is not a measurement on every board.

Does a 4 gigabyte model prove local safety?

No. A size is a deployment fact. A safety property would be a test. The page this brief uses is not that test.