DeepSeek · model № 005
DeepSeek-V4.1-Flash
A 552B model that spends 8B parameters on your prompt and 16B on each new token, with a cache that costs 890 bytes per token
backbone params
552B
+196B in an Engram table
active / token
8B / 16B
prefill / decode
global KV / token
890 B
1/4 of V4-Flash
price in / out
$0.15/$0.60
cache reads at $0.003
Chapters
- 1The Input BillAgent workloads are mostly input, and input is the part of a request this model rebuilt itself to process cheaply.
- 2Two Models in OneA causal encoder builds the cache, a decoder reads it. Half the layers stop running during prefill, and the bill follows.
- 3Three Modes, One CacheLayers take turns building one shared cache, then reuse it with fresh selections or no selection at all.
- 4890 BytesWhat one token of remembered history costs in four-bit arithmetic, and how many bytes a million tokens become.
- 5Replay, Don't StoreThe sliding window state no longer waits on a disk. Serving rebuilds it from the last 128 tokens when it is missing.
- 6Four AdditionsVision, a lookup table for memorized patterns, a drafting head for fast decode, and a rewritten residual path.
- 7Reading the ScoreboardWhere V4.1-Flash beats models several times its size, where it does not, and what the cache discount does to a monthly bill.