The recent update adding local model support is intriguing, but my initial tests have been a mixed bag. I appreciate the direction—reducing reliance on a single API provider is a smart move for a tool like Aider. However, "support" and "practical usability" feel like two different things right now.
I've been experimenting with a few open-source models via Ollama. While the integration is technically sound, the experience isn't yet comparable to GPT-4 for complex coding tasks. The models often struggle with:
* **Context window management:** They seem to lose the thread of larger changes more frequently.
* **Precision:** Instructions need to be far more explicit, and the output can require more correction.
* **Project-wide reasoning:** The leap from editing a single file to understanding cross-file dependencies is where they falter most.
My current take is that it's usable for smaller, well-scoped refactors or boilerplate generation if you have a powerful enough local setup. For major features or architectural changes, the iterative back-and-forth cost is still too high compared to the cloud alternatives.
I'm curious about others' practical experiences. Have you found a specific local model and workflow that makes this viable for your day-to-day work? What hardware are you running on, and have you tweaked any parameters to get better results?
—Anita
—Anita
You've pinpointed exactly where the friction is. I've had similar results using Ollama with some of the larger open-source models.
The key for me has been treating it like a junior engineer who needs very clear, incremental tasks. Asking it to "refactor the authentication module" will go poorly, but "convert this single class from using callback patterns to async/await" often works well. It's usable for those discrete, well-bounded changes.
I'm curious if you've found any particular model or prompt pattern that helps with the cross-file dependency issue? That's the real blocker for moving beyond small tasks.
Reviews build trust.