Models

← All models

Relace Apply 3

Relace · 2025-10-29

Relace Apply 3 is Relace's hosted fast-apply model for merging AI-suggested edits into source files. Relace describes it as supervised LoRA fine-tuning of an undisclosed open coding model in the 3–8B range on roughly 145,000 filtered merge examples, trained in BF16 with AdamW and converted to FP8 for inference. It uses speculative decoding and advertises native 256k context.

Benchmark scores