Ollama 0.40.5 brings NVIDIA's Nemotron 3.5 Super and its smaller sibling, Bolt

Ollama 0.40.5 adds two NVIDIA models you can run on your own machine: Nemotron 3.5 Super, a vision-language model for coding, reasoning, tool use and reading documents , and Bolt, a smaller sibling tuned to fit a 24 GB NVIDIA graphics card. The release page carries the date 14 October, 19:45 , but that is the day the page was opened, while the release was still a pre-release; the final build came out on 15 October, when its installers were uploaded at 18:49 UTC. Below is what each model needs from your machine, what else changed, and who should update.

What changed

Super is the large one. It pairs a 120B-parameter mixture-of-experts backbone — a hybrid of Mamba-2 and Transformer layers — with a C-RADIO vision encoder, and only 12B of those parameters work on each token. Mixture of experts means the model is split into many specialist parts and only a few of them run for each piece of text, so it answers faster than its full size suggests; but every part still has to sit in memory. "Vision-language" means it reads images as well as text — screenshots, charts, scanned pages. It also calls tools, has a thinking mode you can turn on or off, and takes a long context. The page's command isollama run nemotron-3.5-super. NVIDIA publishes it under the OpenMDW License Agreement, version 1.1.

Bolt is Super distilled down to 74B parameters, 9B of them active, and tuned to fit a 24 GB NVIDIA card for everyday agentic tasks — work in which the model plans steps and calls tools. It speeds itself up with speculative decoding through a drafter that comes with it : a small model guesses the next few words and Bolt checks the guess, keeping what it agrees with. Its file squeezes the expert weights into the very compact IQ2_XXS format, trained for that squeeze, and stores the other layers in NVFP4 and Q8_0. That is how 74B parameters fit in a download this small. You run it withollama run nemotron-3.5-bolt.

Both models need the engines underneath to know them: this release updates llama.cpp, adding Nemotron 3.5 support, and MLX.

The rest of the list:

• Decision models now run in Ollama's cloud: clef:cloud and clef-flash:cloudrun them on ollama.com without downloading them to your machine. A decision model answers with a choice, a probability or a score instead of sentences; our itemOllama 0.35 adds decision models that answer with choices, probabilities and scores instead of text explains them.
• Continuing an OpenAI-compatible Responses conversation that includes custom tool calls no longer fails with an error.
• When cloud access fails, the error now says why: a sign-in problem, a payment issue in the desktop app, or an upstream access error from the ChatGPT gateway.

What it was before

The page lists its changes as those since v0.40.4. That release brought decision models to the terminal; our itemOllama 0.40.4 lets you ask decision models questions straight from the terminalcovers it. Until now those models ran only on your machine; the cloud tags are what this release adds.

What it means for you

Start with the download size, because it decides which of the two you can hold. Ollama's library lists Super at 80GB - 90GB and Bolt at 25GB. A model runs well only when its file fits in the memory of the part doing the work — a PC's graphics card, or a Mac's unified memory — and our guideWhich Machine Runs a Model Locally explains that arithmetic on the machine you already have.

Super's 80 to 90 GB is more than three times the memory of a 24 GB card, so no such card holds it; it is for machines with that much memory to spare. The space beyond the file matters too: the library page says your memory and settings decide how much of the long context you can actually use, and that your prompts, images, history, tool results and the model's own output all share that budget. If you can just about fit the file, keep conversations short.

Bolt is the one aimed at a 24 GB NVIDIA card. Two figures sit close together here: the card it is tuned for has 24 GB, and the library lists the download at 25GB. The page does not explain the gap, and it says memory needs depend on how much context is in use, on images and on speculative decoding. On a 24 GB card, begin with short text-only prompts and add images and longer conversations once you see how it runs. NVIDIA provides CUDA instructions for it , and Apple Silicon is not currently supported — on a Mac, Bolt is not your model for now; our guideMac or PC for Running Models Locally explains why the two machines differ. A card with less than 24 GB is below what Bolt was tuned for.

If you call Ollama through its OpenAI-compatible Responses interface with custom tools, update: continuing such a conversation no longer stops with an error. If you want a decision model but lack the memory, the cloud tags let you try one without a download — the text you send is then handled on ollama.com, not on your machine.

The source

The release page on GitHub, Ollama v0.40.5, and its list of downloadable files with their upload times; Ollama's library pages for nemotron-3.5-super and nemotron-3.5-bolt.