A Year of Fast Apply: The Path to 10k Tokens per Second
relace.ai
Relace open-sourced its training and evaluation approach for Fast Apply models after releasing its first public Fast Apply model a year ago and improving code-merge inference. Relace Apply 3 is designed to run at 10k+ tokens per second after generating mergeable diffs and applying them with a specialized model. This changes how coding edits are produced and merged by splitting diff generation from a fast apply step, using an LLM-as-a-judge pipeline to train a smaller merge model on ~145k filtered examples.
Why it matters
Relace explains how it built a small model that applies code edits at 10,000 tokens per second using high-quality merge data, LoRA fine-tuning, FP8 quantization, and speculative decoding. Its training pipeline narrowed 200,000 candidates down to about 145,000 with a judge calibrated against 500 manually reviewed merges.