A system successfully prefills a 284B parameter DeepSeek-V4-Flash model on NVIDIA DGX hardware and decodes it on Apple Silicon Mac Studio over standard 10GbE ethernet, achieving 1.5x to 3.7x speedup on prompts up to 241K tokens by computing the decoder's cache on the prefill machine rather than transferring incompatible KV cache formats.