火山场景构建模型实测对比

Old DeepSeek V4.1 Flash vs the new V4.1 Flash 1017, head to head on 2 DGX Sparks. Same prompt, same settings, thinking on. Each model built a volcano, then got up to 3 rounds to review its own page. Each round it saw real screenshots plus a written description of what was on screen. The reviewer only stated facts and never suggested fixes. The model alone decided: KEEP, or edit the file. No human intervention. -Original: first try 66.6K tokens / 29 min. It said KEEP in round 1: it reasoned the eruption is timed to start at 29s, so the early screenshots just hadn't reached it yet. It was right. Total: 85.8K tokens / 39 min -1017: first try 56.1K / 25 min, then fixed it 3 times (53.7K, 29.2K, 36.0K). It never said KEEP; each version erupted sooner and bigger. Round 3 was its last chance. Total: 175K tokens / 75 min 1017 is cheaper per answer and much stronger on long builds, but its judgment is less steady. Edited the volcano 3 times even though it was already working