The project says Blackwell kernels are compatible with Rubin and that Rubin-tuned attention, GEMM, and MoE kernels are available through FlashInfer 0.7.0.
The preview describes use of CUDA 13.4 locality domains to place MoE weight shards near the compute units that consume them. vLLM reports an early result of 7.8× throughput per GPU versus GB200 NVL72 on AgentX at matched interactivity. The post characterizes the result as early; it should not be read as a general performance guarantee. https://vllm.ai/blog/2026-10-09-vera-rubin-preview