antirez’s DS4 Runs DeepSeek V4 Flash at 39 Tok/s [2026]

An Italian programmer has achieved a striking milestone: running DeepSeek's latest AI model at speeds that rival commercial deployments on consumer hardware, raising fresh questions about whether the era of expensive cloud-based AI infrastructure is nearing an end. The demonstration, which processes language at 39 tokens per second on modest equipment, suggests that frontier AI capabilities are distributing far faster than major tech companies anticipated, potentially upending the economics of their multibillion-dollar investments in centralized data centers. As open-source models accelerate and hardware efficiency improves, the competitive advantage of closed ecosystems like OpenAI and Google may depend less on raw capability and more on the services and scale they can wrap around it.

shattered.io·Sep 3