DeepSeek v4.1-flash tested: agentic coding jumps to flagship-tier, agent behavior turns conservative
karminski3 · x · 2026-10-20
karminski3 tested deepseek-v4.1-flash-1017: backend agentic coding jumped from roughly GROK-4.6 level to domestic flagship level, raising doubts about whether v4.1-pro is still needed. Frontend skills were heavily post-trained, producing flashy three.js demos. Agent behavior shifted a lot: the previous version was "proactive," the new one more "conservative" and harness-dependent; budget control seems worse but improves when budgets are emphasized in-session. A detailed video is expected this week.
Related event: DeepSeek V4.1 Flash Tested: Agentic Coding Jumps to Domestic Flagship Level(2 posts)→
More from coding & agent
- Dev ports Minecraft 26.3 terrain noise to GPU shader, hitting 280 FPS with 32 draw calls in VRChat — Michael_Moroz_ · 2026-10-20
- CoworkEnv synthesizes executable training environments for long-horizon cowork agents — Zhongwei Xie · 2026-10-20
- UCLA and Apple's OpenVLHarness adds multimodal memory to visual agentic tool-use — Yu Zhou · 2026-10-20
- Interlock: Typed Completion Graphs Push Long-Horizon Agents from Local Correctness to Task Completion — meituan-longcat · 2026-10-20
- TokenWorth: Nanjing Univ. Teaches Lifelong Agents Which Memories Are Worth Their Token Cost — nanjinguniv · 2026-10-20
- HarnessPro: Label-Free Recursive Harness Self-Improvement Balancing Agent Performance and Cost — Zhuoshi Pan · 2026-10-20