DeepSeek V4.1 Flash Speeds Increase on GB300 NVL72

Speeds for DeepSeek V4.1 Flash on GB300 NVL72 have increased by up to 1.5 times tokens per second per agent over the past month on the AA-AgentPerf benchmark. The performance improvements stem from a newer vLLM version, new vLLM kernels for DeepSeek V4.1 on GB300, FlashInfer autotuning, fuller CUDA graph coverage, tuned speculative decoding, and running prefill and decode on the same GPUs in NVIDIA's updated Dynamo and vLLM setup. Open article →