Epilude has released Model 4, the latest iteration of its on-device dictation model powering Local Mode. This update represents a significant quality leap, offering improved writing accuracy without compromising privacy or speed. Model 4's cleanup model, built on the Qwen family, refines raw speech-to-text into polished writing, handling punctuation, filler words, and tone adjustments.

Unlike average-based metrics, Epilude grades models on "critical errors" – those that alter meaning or reveal instructions. Model 4 adheres to a strict shipping rule, introducing no new critical errors compared to its predecessor. On Epilude's internal benchmark, Model 4 surpasses Model 3, with a 56 percent reduction in critical meaning errors.

Key improvements are seen in handling self-corrections, dates, mixed-language speech, and casual tone. The model averages multiple training runs to enhance performance and mitigate individual quirks. Model 4 requires Apple Silicon and 16 GB of memory for full functionality. It is available at no extra cost to Epilude users prioritizing privacy, as all processing occurs locally on the Mac.