Trying to find a decent solution for local AI coding has been difficult for me. My GPU is only 12 GB, although it is Nvidia. So I am relegated to smaller models or Mixture-of-Experts models. I have had some pretty bad experiences with performance and crazy file editing problems.
After taking a little AI break, I decided to use Laguna XS 2.1 and refactor (simplify) a Rust function. I had written this function by hand and I wanted to clean it up. I had over-optimized it, and it turns out a different optimization makes this function less important for performance than I had initially assumed.
With OpenCode and Laguna XS 2.1:
• I renamed a function, changed how it tracks results, and changed how it iterates over possible matches.
• The harness and LLM worked well with minimal errors, and my task was successfully completed.
With small local models like Laguna XS 2.1, the trick is to treat them as a typing assistant. You have to tell the model what to do—what arguments to add and remove. It only works for minimal steps at once. So it is mostly like a typing accelerator—not a fully automatic vibe coding machine.