A coding agent for very small models.

sprint runs in your terminal and drives models of 0.5B to 3B parameters on your own machine, inside an 8K context. Two Python files, no dependencies.

pip install git+https://github.com/Vasiniks/sprint.git

Then run sprint. Needs Python 3.9 or newer and Ollama.

One message, start to finish

In the default mode, sprint shows each command and each file change before it happens.

Fast, or right more often

Two models that fit in under a gigabyte. Each row fills at that model's speed; a solid cell is a task it finished.

Qwen 3.5 0.8B122 tokens a second, passed 25 of 48
magi 1.2 2B75 tokens a second, passed 42 of 48

Six one- and two-step tasks, eight tries each, on one Mac. Results swing with wording, so treat them as rough.

Built around what small models get wrong

They lose track of long conversations

sprint keeps room free for the reply and folds older steps into a short summary of what you asked and what was done. File changes outlast things that were only looked at.

They go in circles

The same call again, the same file rewritten, the same failure twice running: sprint warns the model once, then stops the turn and tells you why.

They garble commands

Every tool call is checked, and a reply cut off halfway is never run. You approve each command and edit, or allow the ones you trust for good. Esc stops a turn at once.

Start

  1. Install Ollama. You don't need to start it; sprint does.
  2. Install sprint: pip install git+https://github.com/Vasiniks/sprint.git
  3. Run sprint in a project folder. The first time, it offers to pull Qwen 3.5 0.8B, a 737 MB download.
  4. Type a request and press Enter. Tab switches between asking first, accepting edits, and running everything.

Works on macOS. Linux and Windows are written and tested in CI but not yet tried by hand. The README has every command and setting.