DeepSeek V4.1-Flash-1017 gets a hands-on agent test
A Juejin AI article reviews DeepSeek's October 17 release of V4.1-Flash-1017, highlighting reported gains in Terminal-Bench, Automation-Bench and ProgramBench, plus vendor claims of lower task costs through token and agent-workflow optimization. The author reproduced DeepSeek's “wish-driven programming” racing-game example through DeepSeek Harness, finding that a one-line request produced a playable game with basic driving, tracks and timers, while a more detailed prompt improved its visuals and polish. The reported second run took 51 minutes and 14 seconds, consumed 60.1 million tokens alongside another smaller task and cost about 3 yuan.
The hands-on reproduction provides practical evidence for DeepSeek V4.1-Flash-1017's agentic coding claims, while the benchmark and cost-efficiency figures remain vendor-reported and should be treated cautiously.
内容编写于 2026.10.25
DeepSeek-V4.1-Flash-1017 发布有一段时间了,虽然已经作为3大主力模型之一用了很久,但一直没有机会认真研究。
尤其是 DeepSeek 官方提出的"许愿式编程(Wish-Driven Programming)"和几个案例,都是相当有看头呀。
DeepSeek-V4.1-Flash-1017,到底升级了什么?
先看官方给出的数据。
10 月 17 日,DeepSeek 发布了 V4.1-Flash-1017 更新。
#DeepSeek#DeepSeek V4.1-Flash-1017#DeepSeek Harness#WorkBuddy#CodeBuddy#Tencent#Three.js