← sources

DeepSeek-V4.1-Flash-1017 on AA & Vals index

source: reddit (reddit:LocalLLaMA)published 2026-10-19 05:56ingested 2026-10-20 02:40

retained text

DeepSeek-V4.1-Flash-1017's Terminal bench 4.0 score:

\- Model card reported 42.3%

\- AA reports 34%

\- Vals index reports 31.82%

[AA](https://preview.redd.it/lvpe9muo5dwh1.png?width=1033&format=png&auto=webp&s=0cb5ca5ff222df8cb1972da3ece128c11f33da4f)

[Vals Index](https://preview.redd.it/rxa3c7hr5dwh1.png?width=1932&format=png&auto=webp&s=ee7d699a3799c30923e818b91edede73d408c199)

[Model card](https://preview.redd.it/g4wz00gt5dwh1.png?width=1030&format=png&auto=webp&s=2787cb371bad96924911513878d0c5db3f60fe2e)

Model card quotes:

>*We use the Claude Code harness for Terminal-Bench 4.0*

I am guessing AA & Vals index uses **mini-swe-agent** thus causing a lower score, also each Terminal bench 4.0 task has a constraint of 8 hours, so box powering the agent & provider's tokens/sec during benchmark can skew results.

Fable 5 score on Terminal bench 4.0:

\- AA reports 42%

\- Vals reports 41.41%

If we take the score of 42.3 from model card, that puts DSv4.1 Flash-1017 on par with Fable 5 (released on June 9, 2026) 

document fetched 2026-10-20 00:42 · canonical https://reddit.com/r/LocalLLaMA/comments/1x9owju/deepseekv41flash1017_on_aa_vals_index