SWE-grep

SWE-grep uses multi-turn reinforcement learning for repository exploration. It can make up to eight parallel tool calls per turn across a four-turn search.

What the release provides

Its tool set uses grep, read and glob operations. Three turns explore the repository; the fourth returns file names and line ranges. This structured output supplies context to the main coding agent.Cognition: SWE-grep and SWE-grep-mini

How to use it well

Practical guidance. Use the returned locations to inspect original code. Search results can omit a dependency or call site, so confirm the relevant paths before changing shared behavior.

Specifications & access

SWE-grep specifications
DetailValue and conditions
AccessProvider product
Product usePowers Fast Context in Devin DesktopDevin: Fast Context

Related models

  • A smaller code-search model used by Fast Context.

    Launch date not confirmedProvider product