Mercury Edit 2

Diffusion LLM that predicts your next code edit in parallel — not word by word

Mercury Edit 2 is the second-generation coding model from Inception Labs, built on a fundamentally different architecture than every major LLM you're used to: a diffusion language model. Rather than generating tokens one at a time in a left-to-right sequence, Mercury operates in parallel — refining a full draft across all positions simultaneously. The result is next-edit prediction that runs up to 10x faster than GPT-4o and Claude 3.5 Sonnet at equivalent quality, with latency that finally matches how fast a human developer types. The model is purpose-built for the "edit" step in agentic coding loops — where an agent needs to predict what change should happen at a given location in a codebase, not generate a full file from scratch. Mercury Edit 2 takes in a code context, a cursor position, and optionally a natural-language intent, and outputs the predicted edit. Benchmarks show it matching or exceeding autoregressive models on HumanEval and MBPP tasks while cutting time-to-first-token by 80%. Inception Labs was founded by researchers from Stanford, UCLA, Google DeepMind, and OpenAI who bet that diffusion would eventually outpace transformers for text the same way it overtook GANs for images. Mercury Edit 2 is the clearest signal yet that this thesis has legs. At $0.25/1M input and $0.75/1M output tokens, it's meaningfully cheaper than GPT-4o-class models — and the speed advantage makes it a natural fit for high-frequency agentic tasks.

Panel Reviews

The Builder

Developer Perspective

Ship

“The speed argument is real — I've integrated it into a Cursor-style flow and the round-trip latency for edits dropped to something that genuinely feels instantaneous. The architecture also means it's less prone to 'over-generating' — it just predicts the edit, not a rambling block of new code.”

The Skeptic

Reality Check

Skip

“Diffusion LLMs have been 'about to beat transformers' for two years. Mercury Edit 2 is faster, sure — but for complex multi-file refactors it still struggles with global context. The benchmark cherry-picking on HumanEval is a red flag when most real coding tasks are messier than a LeetCode problem.”

The Futurist

Big Picture

Ship

“This is the first credible sign that the transformer monoculture in language AI might actually break. If diffusion models hit parity on reasoning while maintaining 10x speed, the cost curve for agentic loops changes completely — and Inception Labs has a year head start on everyone else.”

The Creator

Content & Design

Ship

“For code-to-design workflows where I'm iterating on UI components in tight loops, the latency improvement is huge. Faster edit prediction means the feedback cycle between idea and implementation collapses — and that changes the creative dynamic substantially.”