How much DeepSeek V4.1 Flash reads and writes

DeepSeek V4.1 Flash reads up to a million tokens and, on Wild West API, writes up to 393,216 of them in a single reply, the longest output of any model on the line.

The two numbers

The window is shared: the prompt and the reply have to fit in it together. A 700,000-token prompt leaves about 348,000 tokens for the answer, whatever max_tokens asks for. DeepSeek's model card recommends a max_tokens of at least 256K, because the reasoning is written into the same budget as the answer.

Why a million tokens is practical here

Long context is usually limited by what it costs to serve rather than what the model can read. DeepSeek says V4.1 Flash needs a quarter of the GPU memory and an eighth of the SSD storage for its KV cache compared with the previous generation (announcement), which is the part of serving that grows with context. Cached input is billed at its own lower rate, so an agent that re-sends the same history every turn pays the cached price for the repeated part.

Against the rest of the line

Every model reads about a million tokens. The difference is how much each writes back in one go: DeepSeek V4.1 Flash's longest reply is three times GLM 5.3's and twelve times GLM 5.3 Flash's, which matters for long reports, large generated files and reasoning-heavy answers.

Long replies and spend caps

A large max_tokens does not mean a large bill. Wild West API prices the worst case before the request leaves, holds it against the key's cap, and lowers max_tokens to what the hold covers, then charges only what was written. See max_tokens clamping and per-key spend caps.

FAQ

Related