Introducing Mercury Edit 2 — Inception Labs
URL: https://www.inceptionlabs.ai/blog/introducing-mercury-edit-2
Author: Kenan Hasanaliyev, Reece Shuttleworth (Inception)
Published: March 30, 2026
Summary
Mercury Edit 2 is a diffusion-based LLM (dLLM) for next-edit prediction in coding workflows. Unlike autoregressive models that generate tokens sequentially, it generates edit predictions in parallel — making suggestions fast enough to feel like “part of your own thinking.” Fine-tuned with KTO (an unpaired RL method) on human preference data from accept/reject feedback.
Key Claims
- Diffusion for speed: dLLMs generate tokens in parallel, enabling faster next-edit prediction than autoregressive models. Specifically optimized for latency-sensitive use (appears in Zed editor).
- Training pipeline:
- Curated edit dataset across languages and scenarios
- KTO (Kahneman-Tversky Optimization) — unpaired RL method using explicit accept/reject feedback
- Result: 48% more edits accepted, 27% fewer edits shown (more targeted)
- KTO over RLHF: KTO works without paired preference data (A > B comparisons). Just needs binary feedback — accepted or rejected. Lower annotation cost.
- Benchmarks: Evaluated on Instinct, FIM (fill-in-middle), NEP (next-edit prediction), and internal benchmark. FIM uses test case execution; others use LLM-as-a-judge.
- Pricing: 0.75/M output tokens. Available via Inception Platform, integrated with Zed editor and ProxyAI.
- Cursor for Composer-2: Mentions similar RL training at Fireworks — showing the RL-from-human-feedback for code tools is becoming standard.
Significance
This is a rare look at a production-grade RL training pipeline for a coding tool — not a research demo. KTO as an unpaired RL method is particularly interesting for settings where paired comparisons are expensive.