Context caching
Understand cache reuse and measured cache-hit pricing.
此链接保留对应的中文文档路径。目前技术参考内容为英文；功能可用性以 DarkSeek 的实际部署和验证状态为准。
Physical cache#
The serving runtime may reuse compatible prompt prefixes in GPU KV cache. This is a runtime scheduling optimization, not durable storage or an account entitlement. Reuse depends on model, prefix, pressure and placement.
Accounting boundary#
The published Models & pricing schedule distinguishes cache-hit and cache-miss input tokens. Do not infer cache-hit token counts from repeated prompts or faster latency. Applying the cache-hit rate requires measured upstream cache usage and reconciled accounting. No cache hit rate is guaranteed, and automated billing is not enabled yet.