DeepSeek V4.1 Flash Tops AA-AgentPerf: NVIDIA GB300 Achieves 261M Tokens/s/MW Efficiency Breakthrough

Artificial Analysis added DeepSeek V4.1 Flash to its agentic inference benchmark AA-AgentPerf, revealing that on NVIDIA GB300 hardware, the model achieves up to 261 million tokens per second per megawatt of power. With roughly one-third the parameters of its predecessor, V4.1 Flash outperforms DeepSeek V4 Pro 0813 in intelligence while demonstrating significant efficiency gains under realistic coding agent workloads. These results provide critical benchmarks for evaluating deployment costs and performance of open-weight models.