SWE-Grep and SWE-Grep-Mini: RL for Fast Multi-Turn Context Retrieval
a day ago
- #AI
- #coding-agents
- #context-retrieval
- SWE-grep and SWE-grep-mini are fast agentic models specialized in highly parallel context retrieval.
- These models match the retrieval capabilities of frontier coding models while being significantly faster.
- Context retrieval in coding agents traditionally faces a tradeoff between speed and intelligence.
- SWE-grep models use up to 8 parallel tool calls per turn, reducing search time from 6 to 4 turns.
- The models are trained with multi-turn reinforcement learning and distilled for efficiency.
- Fast Context subagent is integrated into Windsurf, reducing time to understand large codebases.
- Evaluation shows SWE-grep models outperform frontier models in speed and retrieval accuracy.
- The goal is to keep users in flow by minimizing response times below 5 seconds.
- Future directions include higher intelligence, variable turn length, and tool speed optimizations.