# DeepSeek flash max is the first model to reach >1% on challenging SWE-sweep bug finding benchmark — retained record

> Retained reddit record from reddit:LocalLLaMA.

Canonical: https://highsignal.app/data/records/5778f086cb825dfb

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:LocalLLaMA)  published 2026-11-13 16:31 ingested 2026-11-14 02:42

[original source ↗](https://www.reddit.com/r/LocalLLaMA/comments/1xvcgxq/deepseek_flash_max_is_the_first_model_to_reach_1/)

## retained text



```
Hi, I've just updated the SWE-sweep leaderboard with more open-weight models! 

DeepSeek v4.1 flash (max) is the first open-weight model we evaluated that gets >1%.

This benchmark is extremely challenging: Models get a snapshot of a (potentially very large) repo containing hundreds of bugs. They don't receive any hint about what these bugs are and need to fix them all. This makes for a very reasoning-heavy, long-horizon task.

To isolate just the model capabilities we also evaluate all models in the same harness (but we did do some comparisons with Claude Code and Codex harnesses and neither seemed to help significantly). However, it is possibly to scale things up somewhat with multiagents. You can run the eval yourself (code on [github](https://github.com/facebookresearch/swe-sweep))

The best proprietary model at the moment is Gemini 4 Hydrogen with 13.5% (according to [google's model card](https://deepmind.google/models/model-cards/gemini-4-hydrogen/)), followed by GPT Astra 6 with 9.5% ([evaluated by us](https://swesweep.com/)). So open-weight models still have a long way to go.

Still extending the open-source leaderboard. What other models should we add? 


```



document fetched 2026-11-14 00:46 · canonical https://reddit.com/r/LocalLLaMA/comments/1xvcgxq/deepseek_flash_max_is_the_first_model_to_reach_1
