Open Models Almanac
Methodsdata curationtokenization

specific method · filed under data curation

Quick Instruction tokens

Quick Instruction tokens are dedicated special tokens appended to the input sequence, with each token corresponding to a specific auxiliary task run alongside response generation.1 By reusing the already-computed KV cache instead of a separate small model, they avoid redundant prefilling and can execute auxiliary tasks in parallel.2

Trade-off It significantly reduces user-perceived time-to-first-token and removes the overhead of maintaining an extra small model.3

How this definition was written

Summarised by glm-5.3-flash from 1 source; every sentence is checked against the quote it cites.

  1. We append a set of dedicated special tokens directly to the input sequence, where each token corresponds to a specific auxiliary task — DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence
  2. By directly reusing the already-computed KV cache, this mechanism completely avoids redundant prefilling and allows certain tasks, such as generating search queries and determining authority and domain, to be executed in parallel — DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence
  3. this approach significantly reduces the user-perceived time-to-first-token (TTFT) and eliminates the engineering overhead of maintaining and iterating an extra small model — DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence

Also called Quick Instruction.

sources
2
models
2
lab adopt it
1
strongest
used

Documented in

Evidence

2 quotes from 2 sources, one card per model, each opening with its strongest.

How sources treat it
used 2

DeepSeek-V4-Pro

DeepSeekadopted2 sources

We append a set of dedicated special tokens directly to the input sequence, where each token corresponds to a specific auxiliary task.

usedinference servingin DeepSeek-V4DeepSeek
1 more from this model

V4 also introduces 'Quick Instruction' tokens: special tokens appended to input sequences that trigger auxiliary tasks in parallel

usedinference servingin DeepSeek-V4DeepSeek
↑