Reasoning

Models that expose their reasoning (OpenAI gpt-5 and o-series, Anthropic extended thinking, Mistral Magistral, Google Gemini 2.5, DeepSeek-R1, local models served by Ollama or vLLM) all surface it the same way in Gwenflow: as ThinkingContent parts, read with response.thinking, and as thinking events on a stream.

Turning it on

Each provider has its own switch:

from gwenflow import ChatAnthropic, ChatGoogle, ChatOpenAI
from gwenflow.llms.openai_response import ResponseOpenAI

# OpenAI, Chat Completions: how hard the model thinks
llm = ChatOpenAI(model="gpt-5-mini", reasoning_effort="medium")

# OpenAI, Responses API: effort and a summary of the reasoning
llm = ResponseOpenAI(model="gpt-5-mini", reasoning_effort="medium", reasoning_summary="auto")

# Anthropic extended thinking
llm = ChatAnthropic(model="claude-opus-4-5", thinking={"type": "enabled", "budget_tokens": 1024})

# Google Gemini 2.5
llm = ChatGoogle(model="gemini-2.5-flash", thinking={"include_thoughts": True, "thinking_budget": 1024})

DeepSeek, Mistral reasoning models and servers started with a reasoning parser return their reasoning without any option.

Reading it

response = llm.invoke("If a train leaves Paris at 9:00 going 120 km/h, when does it reach Lyon, 465 km away?")
print("Reasoning:", response.thinking)
print("Answer:", response.text)

On an agent, AgentResponse.reasoning_content gathers the reasoning of every model call of the run, and a stream yields it as AgentEventThinking events:

for event in agent.run_stream("..."):
    if getattr(event, "event_type", None) == "thinking":
        print(f"[thinking] {event.content}")

Inline `<think>` tags

Local models such as Qwen3, DeepSeek-R1 distills or Gemma 3 often write their reasoning inline as <think>...</think>. ChatOllama extracts these blocks into thinking parts automatically; for another OpenAI-compatible server, set extract_think_tags=True on ChatOpenAI.

Thinking and tools

With Anthropic extended thinking, the reasoning blocks of a turn that called tools must be sent back verbatim, signature included, on the next request. The agent does this for you: the thinking parts are kept on the assistant message (Message.thinking_parts) and replayed. Other providers ignore them.

A separate reasoning model

Agent(reasoning_model=...) asks a second model to work through the task first. Its conclusions are added to the history as an assistant message before the main model takes over, its usage is counted in the run, and its own reasoning (when the model exposes it) is reported in reasoning_content:

agent = Agent(
    name="Planner",
    llm=ChatOpenAI(model="gpt-5-mini"),
    reasoning_model=ChatOpenAI(model="o4-mini"),
)

Get started

Concepts

Models

Agents

Multi-agent

Extend

Memory

Flows

Knowledge

Production

Project

Turning it on

Reading it

Inline `<think>` tags

Thinking and tools

A separate reasoning model

​Turning it on

​Reading it

​Inline <think> tags

​Thinking and tools

​A separate reasoning model

Turning it on

Reading it

Inline `<think>` tags

Thinking and tools

A separate reasoning model