Ollama has added a compact AI model that runs directly on a computer and handles images and documents. Developers can test reading visual material and making repeated tool calls with a smaller download than the larger version.

1. A compact model for images and tools joins the lineup

Ollama v0.40.10, released on October 23, introduces NVIDIA's Nemotron 3.5 Bolt E4B. The release notes highlight fast local chat, image understanding, and tool use. Context, the information a model can refer to in a chat, is 128K.

According to the Ollama library, NVIDIA further trained this model on Gemma 4 E4B. It has 8 billion total parameters, values adjusted during training, with 4 billion effective parameters. It targets instruction following and repeated tool use.

The addition lets users try image-related tasks with the same local model, beyond answering text questions. The library says it can handle images, screenshots, charts, and documents.

2. Connecting visual inspection with tool calls

Image understanding uses screens or documents as material for questions. Tool use calls needed functions and uses their results in the next task. Developers can test reading visuals and continuing work with tool results together.

The library also describes an option to turn reasoning on or off before answering. Reasoning is the internal computation used to produce an answer. Users can choose whether to use it for their purpose.

A DFlash draft model is also provided to assist generation alongside the main model. Citing the company's announcement, the library says it speeds up generation through speculative decoding, using drafts to advance text generation.

3. Smaller downloads come with a shorter context

The library's model list gives E4B a download size of 7.2GB–7.9GB. The larger base version in the same list is 25GB. Smaller files to download and keep can reduce storage requirements when adopting a local model.

E4B has a 128K context, versus 256K for the larger base version. This offers a comparison between smaller files and longer context. Users extending chats and tool results need to consider context as well as download size.

Usable context depends on memory, with images, chat history, tool results, and output sharing it. Neither source provides comparative speed or quality measurements, so the acceleration claim cannot yield a specific performance difference.

4. Run the smaller version by specifying its tag

The library lists support for Windows, Linux, and Apple Silicon for E4B. Run it in a terminal with the following command from the release notes.

ollama run nemotron-3.5-bolt:e4b

The final e4b tag, a label distinguishing model versions, selects the smaller version. In the library list, latest points to the 25GB base model. To try the smaller model, specify it as shown in the provided command.

The library page also includes Python and JavaScript chat call examples shared across the model family. These offer entry points for connecting programmatic chat features, beyond chatting in a terminal.

Guides also show how to launch coding tools such as Claude Code and OpenCode with a specified model. This release focuses on a smaller model for local image and tool tasks rather than new calling conventions.

E4B expands the options for testing local image processing and tool use, with a trade-off between download size and context.