Thinking

When the request context contains only a <Think> tag, this is enough for the LLM to understand that it must start producing tokens which look like thinking and not look like final answer.

how thinking stops?

The thinking token generation stops whenever the model predicts </Think> as the next token during the decode process.

How thinking is trained into LLMs?

During the training process, LLMs are trained with data which actually has the thinking tags, thinking/reasoning tokens and then the final answers.

With such enormous data with thinking pattern, the LLMs learns to generate thinking like tokens before generating the final answer.

how thinking is disabled?

When an agent configured to disable thinking for a model, then the agent simply sends the opening <Think> as well as the closing </Think> token. With this, the LLM will never generate reasoning tokens.

Reasoning Tokens Limit

This is the limit on number of reasoning/thinking tokens per inference request. When the number of thinking tokens cross the configured limit, the inference engine at the end of a full forward pass, checks the token count and forcefully inserts a </Think> tag to stop the thinking/reasoning if it has exceeded the limit.

what happens when there is no limit?

When there is no limit, the LLMs can produce unlimited thinking tokens, which can be useless, expensive and time-consuming.