V4.1-Flash-1017 多项跑分超 Opus 5，552B MoE 权重可下载
DeepSWE 77.0 对 74.0、Automation-Bench 60.5 对 50.3、Agents' Last Exam 35.4 对 28.6、Apex 72.1 对 65.8，但 Terminal-Bench 与 HLE 仍落后；作者强调这是 552B MoE、激活 8B/16B 的最小档模型且权重开源
CHINA IS CLOSING THE FRONTIER AI GAP - AND YOU CAN DOWNLOAD THE MODEL DeepSeek just released V4.1-Flash-1017. And the benchmark results deserve attention. DeepSWE: 77.0 vs Claude Opus 5’s 74.0 Automation-Bench: 60.5 vs 50.3 Agents’ Last Exam: 35.4 vs 28.6 Apex Bench: 72.1 vs 65.8 Opus 5 still leads in other areas, including Terminal-Bench and Humanity’s Last Exam. But here’s the bigger story. DeepSeek is achieving these results with a 552B-parameter Mixture-of-Experts architecture that activates just 8B parameters for input processing and 16B for output generation. And the weights are available to download. Community developers are already running the model locally on NVIDIA DGX Spark systems, using quantization and SSD offloading. Not exactly plug-and-play, not very performative. But increasingly possible. Meanwhile, the world’s most powerful closed models still live behind APIs controlled by a handful of companies. And remember: DeepSeek calls V4.1 Flash the SMALLEST model in its new architecture family. The implications go far beyond another benchmark victory. The AI race is no longer just about who builds the smartest model. It’s also about who makes powerful intelligence accessible to everyone else.