Open Models Almanac
Methodsinference & servingreasoning control

general family · filed under inference & serving

Thinking mode

Thinking mode is one of two reasoning modes of a single model, the mode in which the model reasons before answering.1 It is selected by choosing the corresponding API model name, e.g. deepseek-reasoner rather than deepseek-chat.2

Trade-off It trades slower answers for stronger reasoning.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. Hybrid reasoning architecture: A single model supports both thinking mode and non-thinking mode — Change Log | DeepSeek API Docs
  2. deepseek-reasoner corresponds to its thinking mode — Change Log | DeepSeek API Docs
  3. Compared to DeepSeek-R1-0528, DeepSeek-V3.1-Think provides answers in significantly less time — Change Log | DeepSeek API Docs

Also called think mode, Reasoning Effort, thinking, Thinking mode (default reasoning mode), Thinking/Non-Thinking modes, dual modes (Thinking / Non-Thinking).

sources
49
models
25
labs adopt it
14
strongest
core

Documented in

Evidence

55 quotes from 49 sources, one card per model, each opening with its strongest.

How sources treat it
evaluated 1optional 24used 12default 14core 4

Kolibri-1

Aleph Alphaadopted3 sources

As summarised in Figure 26, the assigned label depends on the length of a reasoning trace relative to the other traces of the same dataset and prompt difficulty.

coredata curationin KolibriAleph Alpha
4 more from this model

Across both stages, we taught the model to reason at different effort levels (none, low, medium, high)

coremodel architecturein KolibriAleph Alpha

You can pass reasoning_effort values low, medium and high to configure the amount of effort our model puts into finding the answer. You can disable thinking altogether by setting reasoning_effort to none

optionalinference servingin KolibriAleph Alpha

Kolibri features three distinct reasoning effort levels, low, medium and high.

coreunclearin KolibriAleph Alpha

training on reasoning traces labelled with an effort level derived from their length relative to requests of the same dataset and difficulty

usedunclearin KolibriAleph Alpha

Gemma 4 31B

Google DeepMindadopted1 source

We introduce a thinking mode [OpenAI, 2024] to Gemma 4 models. By outputting a reasoning trace before the response, models demonstrate improved capabilities in reasoning-heavy domains such as mathematics and coding.

corepost trainingin Gemma 4Google DeepMind

Mistral-Large-4

Mistral AIadopted4 sources

three reasoning efforts — low ,medium ,high (default)

defaultinference servingin Mistral Large 4Mistral AI
4 more from this model

the final model supports high, medium, and low reasoning effort, giving users control over the trade-off between token consumption and performance

optionalpost trainingin Mistral Large 4Mistral AI

Mistral Large 4 can take low, medium, high as reasoning effort. The default is high

optionalinference servingin Mistral Large 4Mistral AI

Mistral Large 4 now supports three reasoning modes

optionalunclearin Mistral Large 4Mistral AI

reasoning effort configurable per request

optionalinference servingin Mistral Large 4Mistral AI

GLM-5.2

Z.ai (Zhipu AI)adopted4 sources

GLM-5 supports controlling the thinking budget through the reasoning_effort parameter, which accepts two levels: max and high. max is the default

defaultinference servingin GLM-5 seriesZ.ai
4 more from this model

GLM-5.2 introduces two selectable effort levels: Max mode and High mode

usedinference servingin GLM-5.2Z.ai

the recommendation of always using the model on Max thinking effort

optionalinference servingin GLM-5.2Z.ai

All models were evaluated in think mode

usedevaluation onlyin GLM-5.2Z.ai

multiple thinking effort levels to balance performance and latency

optionalinference servingin GLM-5.2Z.ai

GLM-5.3

Z.ai (Zhipu AI)adopted3 sources

GLM-5.3 supports controlling the thinking budget through the reasoning_effort parameter, which accepts three levels: low, high, and max. It defaults to max if not passed (or if set to any other value).

defaultunclearin GLM-5.3Z.ai
2 more from this model

GLM-5.3 always operates with reasoning enabled and supports three reasoning effort levels: low, high, and max.

defaultinference servingin GLM-5.3Z.ai

GLM-5.3 offers three reasoning effort levels driven by the reasoning_effort field

defaultinference servingin GLM-5.3Z.ai

Kimi K3

Moonshot AIadopted2 sources

Thinking effort is configured with the top-level `reasoning_effort` request field, which supports "low", "high", and "max" (default "max").

defaultinference servingin Kimi K3Moonshot AI
1 more from this model

At launch, Kimi K3 will use max thinking effort by default, with low- and high-effort modes to be introduced in subsequent updates.

defaultinference servingin Kimi K3Moonshot AI

Qwen3.6-35B-A3B

Alibaba (Qwen)adopted2 sources

Qwen3.6 models operate in thinking mode by default, generating thinking content signified by thinking ... response before producing the final responses.

defaultinference servingin Qwen3.6-35B-A3BQwen Team
1 more from this model

Qwen3.6 will think by default before response.

defaultinference servingin Qwen3.6-27BQwen Team

Laguna-S-2.1

Poolsideadopted1 source

Native reasoning support: You can run requests with thinking off or with max thinking. Max thinking is enabled by default

defaultinference servingin Laguna S 2.1Poolside

gpt-oss-120b

OpenAIadopted5 sources

By default, the model will do medium level reasoning. To control the reasoning you can specify in the system message the reasoning level as low, medium, or high.

optionalinference servingin gpt-ossOpenAI
4 more from this model

Configurable reasoning effort: Easily adjust the reasoning effort (low, medium, high) based on your specific use case and latency needs.

optionalunclearin gpt-oss-120bOpenAI

Easily adjust the reasoning effort (low, medium, high) based on your specific use case and latency needs.

optionalinference servingin gpt-ossOpenAI

We train the models to support three reasoning levels: low, medium, and high.

usedpost trainingin gpt-oss-120b & gpt-oss-20bOpenAI

The model supports three reasoning efforts: low, medium and high.

optionalinference servingin gpt-oss-120b & gpt-oss-20bOpenAI

DeepSeek-V4-Flash-0731

DeepSeekadopted3 sources

The reasoning_effort parameter now supports three levels — low, high, and max — which control how much deliberation the model spends before answering.

usedinference servingin DeepSeek-V4-Flash-0731DeepSeek
2 more from this model

The thinking modes of V4-Pro and V4-Flash now support three thinking effort levels: low / high / max.

optionalinference servingin DeepSeek V4-Pro and DeepSeek V4-FlashDeepSeek

the reasoning effort control was also expanded to three explicit levels (low, high, max)

usedinference servingin DeepSeek-V4-FlashDeepSeek

NVIDIA-Nemotron-3-Ultra-550B-A55B

NVIDIAadopted2 sources

Ultra is also equipped with reasoning effort control, which supports inference-time adjustment of the accuracy–compute trade-off

usedinference servingin Nemotron 3 UltraNVIDIA
3 more from this model

Ultra is also equipped with reasoning effort control, which supports inference-time adjustment of the accuracy–compute trade-off

usedinference servingin Nemotron 3 UltraNVIDIA

The medium-effort reasoning mode is introduced during the SFT stage and later optimized during the RLVR stage.

usedpost trainingin Nemotron 3 UltraNVIDIA

Nemotron 3 Ultra is trained for three reasoning modes: reasoning-off, regular and medium-effort.

optionalpost trainingin Nemotron 3 UltraNVIDIA

MiniMax-M3

MiniMaxadopted3 sources

M3 supports toggling thinking on or off. With thinking enabled, the model is suited to complex reasoning, agentic tasks, and long-horizon collaboration; with it disabled, it responds faster

optionalinference servingin MiniMax-M3MiniMax
2 more from this model

A thinking mode can be toggled on or off per request.

optionalinference servingin M3 APIMiniMax

M3 supports three reasoning modes through the thinking parameter

optionalinference servingin MiniMax-M3MiniMax

DeepSeek-V4-Pro-0813

DeepSeekadopted2 sources

The reasoning_effort parameter now supports three levels — low, high, and max — which control how much deliberation the model spends before answering.

optionalinference servingin DeepSeek-V4-Pro-0813DeepSeek
1 more from this model

The thinking modes of V4-Pro and V4-Flash now support three thinking effort levels: low / high / max.

optionalinference servingin DeepSeek V4-Pro and DeepSeek V4-FlashDeepSeek

Step-3.7-Flash

StepFunadopted2 sources

offers three selectable reasoning levels (low, medium, and high) so developers can easily balance speed, cost, and cognitive depth

optionalinference servingin Step 3.7 FlashStepFun
1 more from this model

Three selectable reasoning depths — low, medium, and high — let developers trade latency for reasoning depth

optionalinference servingin Step 3.7 FlashStepFun

DeepSeek-V3.2

DeepSeekdiscussed1 source

The performance of non-thinking mode is slightly worse than the thinking mode, but still competitive.

evaluatedinference servingin DeepSeek-V3.2DeepSeek

Further reading

Picked by hand, not extracted: where to read more, not evidence for anything on this page.

↑