deepseek-v4.1VIDEO

DeepSeek V4.1 Flash: bare single-call test vs harness agent loop

A user compared DeepSeek V4.1 Flash in a single bare API call (under 100K tokens) against a harness run of 3 rounds and ~400 steps (~90M tokens, 99% cached), finding the bare call already shows model capability.

AI 观察家 Andy

@renmu2017Archived 12 days agoOriginally posted 2026.10.19

AI 观察家 Andy @renmu2017Originally posted 2026.10.19

Open original on X

DeepSeek-v4.1-flash-1017：不接工具，单次调用也表现不错接上 harness 确实能显著改变输出但不跑 Agent 循环也能看出模型本身的能力两种方式用的是同一个提示词 1. 裸测只调用一次 API 不用 harness、工具或循环总 token 数不到 10 万 2. 接入 harness 跑 3 轮、约 400 步用了约 9000 万 token 其中 99% 是缓存输出约 40 万 token Deepseek这次的更新碉堡了！ https://x.com/ItsmeAjayKV/status/2111882446094692823/video/1

x.com/renmu2017/status/2112019383648387112Archived 2026.10.19 12:02