DeepSeek reports 890 bytes per token

DeepSeek-V4.1-Flash reports a global KV-cache footprint of 890 bytes per token, roughly one-quarter of its predecessor, in the model’s technical report. At equal memory capacity, that implies room for about four times as many long-context agent sessions, but the paper’s figure excludes allocator overhead, activations and transient state. Teams comparing DeepSeek’s low-cost vision serving economics should treat this as a serving-density result to benchmark, not a promise of fourfold production throughput.