Open Models Almanac
Methodsinference & servingreasoning control

specific method · filed under inference & serving

Continuously controllable reasoning effort

A continuously controllable reasoning-effort setting is an integer (1–100) control over how much the model reasons at inference time.1 It lets a deployment trade inference cost against accuracy by adjusting the reasoning strength.2

Trade-off Higher reasoning effort buys accuracy at the cost of more inference compute.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. supports a continuously controllable reasoning effort setting (integer 1–100 — deepseek-ai/DeepSeek-V4.1-Flash · Hugging Face
  2. that trades inference cost for accuracy — deepseek-ai/DeepSeek-V4.1-Flash · Hugging Face
  3. trades inference cost for accuracy — deepseek-ai/DeepSeek-V4.1-Flash · Hugging Face

Also called Deployment-time reasoning-effort control, reasoning-effort setting, Scalar effort control of response length, scalar efforts, Scalar reasoning-effort control.

sources
3
models
2
labs adopt it
2
strongest
core

Documented in

Thinking Machines Lab

Evidence

6 quotes from 3 sources, one card per model, each opening with its strongest.

How sources treat it
used 4core 2

DeepSeek-V4.1-Flash

DeepSeekadopted2 sources

The scalar effort variable provides a deployment-time control over the cost–quality trade-off without imposing a hard token budget.

coreinference servingin DeepSeek-V4.1-FlashDeepSeek
4 more from this model

We therefore introduce a scalar effort level 𝑏 as an explicit conditioning signal during reinforcement-learning.

coreunclearin DeepSeek-V4.1-FlashDeepSeek

The model supports a continuously controllable reasoning effort setting (integer 1–100) that trades inference cost for accuracy.

usedinference servingin DeepSeek-V4.1-FlashDeepSeek

By varying 𝑏, a single model checkpoint can move between different operating regimes along the learned cost–quality frontier

usedunclearin DeepSeek-V4.1-FlashDeepSeek

the use of scalar efforts achieves flexible, interpolated control of response length within a specific range.

usedinference servingin DeepSeek-V4.1-FlashDeepSeek

Inkling

Thinking Machines Labadopted1 source

We specified the model’s effort level on different samples by changing the system message and adjusting the per-token cost. This caused the model to use a different amount of tokens in different rollouts and learn the ability to control thinking effort.

usedpost trainingin InklingThinking Machines Lab
↑