# We evaluated DeepSeek V4.1 Flash 1017 and Kimi K3.1 Pro on LongTail-SWE — retained record

> Retained reddit record from reddit:LocalLLaMA.

Canonical: https://highsignal.app/data/records/50bf3538a891b9bd

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


source: [reddit](https://highsignal.app/data/reddit) (reddit:LocalLLaMA)  published 2026-11-18 16:38 ingested 2026-11-19 02:39

[original source ↗](https://www.reddit.com/r/LocalLLaMA/comments/1xzjtu2/we_evaluated_deepseek_v41_flash_1017_and_kimi_k31/)

## retained text



```
Last week we shared LongTail-SWE, 202 real GitHub tasks in 13 less common languages. Many of you asked for DeepSeek V4.1 Flash 1017, so we ran it along with Kimi K3.1 Pro.

DeepSeek V4.1 Flash 1017: 76%, up from 57% for V4 Flash.  
Kimi K3.1 Pro: 77%, the top open-weight model, with 5.6M tokens per attempt.

We also re-ran Kimi K3 with a 1M context instead of 262K (now 72%, +8 pts) and fixed tests in a few tasks.

Leaderboard: [longtail-swe.com](http://longtail-swe.com)
```



document fetched 2026-11-19 00:42 · canonical https://reddit.com/r/LocalLLaMA/comments/1xzjtu2/we_evaluated_deepseek_v41_flash_1017_and_kimi_k31
