Ollama Downloader in Rust (ODIR)

{} :: # # # # # # # #

Ollama is an open-source tool that faciliates the user to download and run large language models (LLMs) such as Llama 3, Mistral, and Gemma on a local computer. It acts as a lightweight runtime and model manager, making it easy to use artificial intelligence (AI) models without relying on cloud services. It does have the capability to run some models on the Ollama cloud too.

A bit of background

While ollama pull <model:tag> mostly works, there is a documented problem, see issue 941. The crux of the problem is that Ollama fails to pull a model with an error message as follows.

Error: digest mismatch, file must be downloaded again: want sha256:1a640cd4d69a5260bcc807a531f82ddb3890ebf49bc2a323e60a9290547135c1, got sha256:5eef5d8ec5ce977b74f91524c0002f9a7adeb61606cdbdad6460e25d58d0f454

This is particularly frequent on corporate networks, which typically have man-in-the-middle (MITM) proxies for HTTPS traffic with self-signed certificates.

My solution in August 2025 was to write a command-line tool in Python that downloads the model file directly from the registry, bypassing Ollama's own method of downloading. This led to the Ollama Downloader project. In February 2026, I wanted to port it in Rust for better performance and support more features, such as parallel downloads and resume capability. This led to the creation of Ollama Downloader in Rust or ODIR, pronounced as oh dear!

More information

You can find the following links related to the ODIR project.