News Decoded · For engineers

Ollama says 128K, NVIDIA says 64K: Nemotron 3.5 Bolt E4B context.

Ollama says Nemotron 3.5 Bolt E4B has 128K of context. NVIDIA's own page says 64K. Here's what we found, and what we'd do before trusting either.

28 Oct 2026 · 6 min read · sources read up to 28 Oct 2026

Seen on Instagram: the post this article goes with.

You planned for 128K, and then the maker said 64K

Ollama says this model has 128K of context ^[1]. NVIDIA's own page says 64K ^[2]. Here's what we found, and what we'd do before trusting either.

Picture this. You're moving an API call over to a local model, and you've sized your prompts around its context window: how many tokens (small chunks of text) it can read at once, where K means a thousand. Then you open the maker's own page and see a smaller number. That's the sinking moment this piece is about: wait, which one is right? We made this scene up, but it's an easy one to walk into, because two official pages really do give two different figures ^[1][2].

Ollama is a free app that runs AI models on your own computer. On 23 October it added NVIDIA's Nemotron 3.5 Bolt E4B in its 0.40.10 release ^[1]. E4B is NVIDIA's small version of the model; the other size is 74B ^[3][10]. Ollama's release note says E4B has a 128K-token context window ^[1]. NVIDIA's own page for the model says input up to 64K tokens and output up to 64K tokens ^[2]. If we'd only read Ollama's note, we'd have planned for 128K too.

We haven't downloaded or run the model ourselves yet, so everything here comes from Ollama's and NVIDIA's own pages and files, and we can't tell you which number holds up in practice. What we can do is show you what each page says, and what we'd do in your place. We'd plan to the lower figure, NVIDIA's 64K ^[2], and then look at what Ollama is really using.

What Ollama's download page lists

In Ollama, a tag is the part of a model's name after the colon, as in nemotron-3.5-bolt:e4b ^[3]. The tags page for this model has ten rows. Three of them belong to a much larger model, which we come back to below. The other seven are the E4B tags ^[3], and here they are next to NVIDIA's figure. The first four disagree with NVIDIA. The last three, Ollama's MLX tags, agree with it.

The e4b and e4b-q4_K_M tags turn out to be the same download, because Ollama's registry files for the two match ^[4][5]. We shortened the names in the Tag column, so the full tag for the second row is e4b-q4_K_M ^[5]. For any other row, copy the exact tag from the tags page ^[3].

The names q4_K_M, q8_0, bf16, mxfp8 and nvfp4 describe how precisely the weights (the numbers that make up the model) are stored, and fuller precision means a bigger file ^[3]. We didn't compare quality between them. This piece is only about the context figure.

The three rows that stopped us

MLX is Apple's framework for running models on Apple chips ^[6]. Ollama's three MLX tags for this model list 64K, the same figure NVIDIA gives ^[3][2]. The other four list 128K ^[3]. So Ollama itself gives two different figures for one model, and that's the part that stopped us while we were reading the pages.

Everything of NVIDIA's we read points the same way. NVIDIA's server commands and its config file both stop at 65536 tokens, which is 64K: the commands on its page for running the model as a server set --max-model-len 65536 ^[2], and the config file for the text model sets its maximum length, max_position_embeddings, to 65536 ^[7]. We searched NVIDIA's page for 128K and 131072 to find where Ollama's figure came from, and found neither ^[2]. None of the pages we read explains the higher figure, so we'd rather not guess at one.

Which context Ollama gives you

For the e4b tag, the context you get comes from Ollama's settings ^[4][8]. Whatever a page lists, ollama ps shows the context your session is really running with ^[8]. You can change it with the slider in the app or with the OLLAMA_CONTEXT_LENGTH environment variable ^[8].

One thing follows from the numbers alone. On a machine with 48 GiB or more of graphics memory, the default on Ollama's context-length page (256K) is above the 64K NVIDIA documents ^[2][8]. That's one more reason to look at what your session is really using before you trust a figure from either page.

Ask for the small model by its tag

One more thing could trip you up. The model's plain name now points to a different, much larger model. nemotron-3.5-bolt:latest is the 74B model (74 billion weights): 25 GB, with a 256K context ^[10][11]. The library page says "Supports up to 256K tokens", but that 256K belongs to the big model ^[10]. To get E4B, ask for its tag: ollama run nemotron-3.5-bolt:e4b ^[1]. The e4b tag needs Ollama 0.40.10 or newer ^[4].

What we'd check before relying on it

1. Pull nemotron-3.5-bolt:e4b by its full tag, so you don't end up with the bigger model ^[1][10].
2. Budget your prompts to the 64K NVIDIA documents, until you've tested a larger figure yourself ^[2].
3. Run ollama ps while the model is loaded, and read the context it shows ^[8].
4. Read the licence yourself: NVIDIA says the model comes under OpenMDW 1.1 and is ready for commercial or non-commercial use ^[2], and as we read it, the text lets you use the model free of charge "without restriction" and asks you to include the licence and notices if you share the weights ^[12].
5. Come back to both pages after each new Ollama release, since the two releases after 0.40.10, on 24 and 27 October, don't mention this model yet ^[13][14].

Our practice: when two official sources disagree, we plan to the lower figure and only move up once we've measured. If the larger figure turns out to be real, that costs some unused room. We'd still rather leave room spare than build around a number the model's maker doesn't document.

What we still can't tell you

We still don't know why Ollama lists 128K on four of its tags, or what the model does past 64K, because nothing we read says ^[2][3].

How we checked

We read every source in the list on 28 October 2026. Besides the web pages, we fetched two kinds of files the vendors publish themselves: Ollama's public registry files for the model (the manifests, plus the config and params files they list) ^[4][5][11] and the config file NVIDIA publishes with it ^[7]. The e4b tag's params file doesn't set a context length ^[4]. Every claim points to the page or file that makes it.

Ollama's own docs don't agree on a default either: 4K, 32K or 256K depending on graphics memory on one page, 2048 on another ^[8][9]. So check ollama ps for your real number.