DeepSeek V4.1 Flash's reasoning_effort scales output quality and tokens ~linearly in tests

zainhas · x · 2026-09-10

Early user testing of DeepSeek-V4.1 Flash's reasoningeffort parameter shows something rare: quality improves nicely and roughly linearly as the setting goes from 1 to 100, and average output token count scales alongside — unlike many models where raising effort yields little.

Original post →

More from Models

Models channel →