Open Models Almanac
Methodsinference & servingcontext management

specific method · filed under inference & serving

Preserved thinking history

Retains reasoning or thinking content in conversation history across turns; the evidence includes configurations that preserve complete assistant messages and reasoning blocks.

The sources on this page name it without explaining how it works. The line above is the extraction's one-line gloss. See context management for the methods filed beside it.

Also called Preserve thinking blocks, preserve_thinking, preserved thinking, Preserved thinking history mode, Thinking Preservation.

sources
9
models
4
labs adopt it
3
strongest
core

Documented in

Evidence

9 quotes from 9 sources, one card per model, each opening with its strongest.

How sources treat it
optional 2used 4default 1core 2

Kimi K3

Moonshot AIadopted4 sources

the think channel is always retained in the history—kept even when its content is empty—so that the model observes a consistent message structure across turns

coresoftware implementationin Kimi K3Moonshot AI
3 more from this model

Kimi K3 was trained in the preserved thinking history mode. For multi-turn conversations and tool calls, Kimi K3 requires the complete assistant message returned by the API to be passed back to `messages` as-is — including `reasoning_content` and `tool_calls`, not just `content`.

coreotherin Kimi K3Moonshot AI

For multi-turn conversations and tool calls, Kimi K3 requires the complete assistant message returned by the API to be passed back to messages as-is — including reasoning_content and tool_calls, not just content

usedinference servingin Kimi K3Moonshot AI

K3 was trained in the preserved thinking history mode.

usedtraining objectivein Kimi K3Moonshot AI

Qwen3.6-35B-A3B

Alibaba (Qwen)adopted3 sources

A new feature retains thinking context across conversation history, streamlining iterative development and reducing overhead.

usedmodel architecturein Qwen3.6Qwen Team
2 more from this model

This capability is particularly beneficial for agent scenarios, where maintaining full reasoning context can enhance decision consistency and, in many cases, reduce overall token consumption by minimizing redundant reasoning.

optionalinference servingin Qwen3.6-27BQwen Team

model, please use"preserve_thinking": Trueinsteadof"chat_template_kwargs": {"preserve_thinking": False}.

optionalinference servingin Qwen3.6-35B-A3BQwen Team

Further reading

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

↑