Sweep AI has released a 1.5 billion parameter model designed for next-edit autocomplete, distinguishing itself from traditional code completion by leveraging recent edits as context. The model runs locally and outperforms larger models in both speed and accuracy.

The team benchmarked their model against tools like Mercury, Zeta, and Instinct, using five different tasks including tab-to-jump and FIM. They found that exact-match accuracy was the most reliable usability metric due to the precision required in code editing.

Training involved supervised fine-tuning on 100k examples and reinforcement learning with tree-sitter parsing checks. Prompt formatting played a larger role than expected, with simple `original`/`updated` blocks outperforming unified diffs.

Sweep is open-sourcing the weights to enable fast, privacy-preserving autocomplete tools across all editors. Developers can experiment with the model via their JetBrains plugin or directly on Hugging Face.