. Filed under reviews and homelab.
Aider with local models through Ollama: two weeks on my own backlog
I pointed a terminal coding assistant at models running on my own GPU and gave it the tickets I would otherwise have done myself. Then I did the same with a hosted model to see what I was giving up.
Sample post written for this site build. The setup is realistic; the task counts, times and costs are illustrative, not measured results.
If you already run a GPU at home, a 32B local model in Aider is good company for small, well-described changes: a new endpoint, a migration, a failing test. It is not yet a replacement for a hosted model on refactors that touch many files. Try it on a week of real tickets before you cancel anything.
What I tested
Aider is a terminal tool that edits files in a git repository from a chat, and commits each change. Ollama serves open-weight models on a local machine behind an API that Aider can talk to. I wanted to know whether the pair could carry routine work without a hosted model.
- 40 tasks from my own backlog across two repositories: a Django API and a TypeScript front end.
- Three setups, each given the same task text and the same files added to the chat.
- A task counted as done when the tests passed and I would have merged the diff without edits.
- Hardware: a home server with one 24 GB consumer GPU, 64 GB of RAM, running Ubuntu.
How the three setups compared
The hosted model finished most tasks unaided. The 32B local model was closer than I expected on single-file changes and fell away once a task needed context from several files. The 8B model was quick to answer and slow to be right.
Tasks done unaided, out of 40
Sample figures. Full breakdown in the table below.
| Setup | Runs on | Done unaided | Needed a fix | Abandoned | Median minutes | Cost | Licence |
|---|---|---|---|---|---|---|---|
| Hosted mid-size model | Provider API | 31 | 6 | 3 | 4 | £30 tokens | Proprietary |
| Local 32B coder, 4-bit | 24 GB GPU | 24 | 9 | 7 | 9 | £6 power | Open weights |
| Local 8B coder, 4-bit | 24 GB GPU | 12 | 10 | 18 | 6 | £4 power | Open weights |
Where it falls down
Context is the limit. On my GPU the 32B model ran comfortably with a modest context window, and tasks that needed four or five files in the chat either ran out of room or slowed to the point where I would rather type the change myself.
The smaller model also edited with confidence when it was wrong. Aider commits each change, so undoing is easy, but every bad commit is still a diff I had to read.
What I liked
- Each change lands as its own git commit with a sensible message, so review is the same as reviewing a colleague’s branch.
- Nothing leaves the machine with the local setups, which matters for client code under a strict contract.
- Switching between local and hosted models is one flag, so you can start local and escalate when a task gets hard.
Who should try it
Developers who already have a capable GPU and a backlog of small, well-scoped tickets. If you would need to buy the GPU for this, the hosted model is cheaper for a long time yet; the monthly market post has the current prices.
Questions or a different result on your own hardware? Email me and I will add it as a note.