Mercury Edit 2

Mercury Edit 2 is the narrow member of the Mercury family: it does not hold conversations and cannot be pointed at a chat endpoint. It serves two developer-tool endpoints instead — fill-in-the-middle autocomplete and next-edit prediction — and its whole reason for existing is that both are the parts of coding where a pause is immediately felt. Reading recent edits and surrounding codebase context, the model guesses what the programmer is about to change next; the user presses Tab to accept. Diffusion is what makes that plausible. Because the model refines tokens in parallel rather than emitting them one by one, a suggestion can land inside the gap between keystrokes instead of arriving after the developer has already typed the line. Inception's own framing, echoed in a quote from Zed co-founder Max Brunsfeld, is that the suggestion should feel like part of the programmer's thinking rather than something waited for. The more interesting engineering detail is that Inception's second version is largely about making the model quieter. The company reports that its first next-edit model was overzealous — too many suggestions, too long — and distracted users. So it recorded explicit accept-and-reject feedback from real use, built a human preference dataset out of it, and aligned the model with an unpaired reinforcement learning method called KTO. The published outcome is a pair of numbers that belong together: edits are accepted 48 per cent more often, and the model is 27 per cent more selective about which edits it shows at all. Restraint, in other words, was treated as a feature and measured as one. Quality is reported as an average over four next-edit benchmarks — the open Instinct, fill-in-the-middle and Next-edit Prediction sets plus one internal suite — of which only the fill-in-the-middle test validates edits by running the modified code; the other three score similarity to a human-written answer using a model as judge, which is a weaker instrument. All results come from the vendor. The model runs through the Inception platform and in editors that support it, Zed among them, at $0.25 per million input tokens, $0.75 output and $0.025 cached input, with 32K of context on both endpoints and up to 8,192 output tokens. Tool calling and streaming are not offered. Inception does not date the posts themselves, but its blog index does: the announcement is listed as published on 30 March 2026. Its status today is split between the maker's own pages: the models page says Mercury 1, Mercury 2 and Mercury Edit 2 remain supported for existing customers only, while the developer documentation still lists Mercury Edit 2 with full pricing and both endpoints.

▸Videos

No videos yet.

▸Reviews