Daily Tokens
November 13, 2026

AI · November 13, 2026 · Low impact

DeepSeek v4.1 Flash (max) becomes the first open-weight model to exceed 1% on the SWE-sweep bug-finding benchmark

The SWE-sweep leaderboard has been updated with more open-weight models, and DeepSeek v4.1 Flash (max) is the first one evaluated to score above 1%. The task gives a model a snapshot of a potentially very large repository containing hundreds of bugs, with no hints, and it must find and fix them all. All models were run in the same harness to isolate model capability, and the benchmark's authors also ran some comparisons using the Claude Code and Codex harnesses.

The benchmark is long-horizon and reasoning-heavy, so an open-weight model reaching even 1% is a concrete data point on how far open models are from frontier coding agents on realistic repositories.

Go deeper

More from AI on Friday, November 13