NInfer - Reasoning Effort

Qwen3.8-27B thinks before it answers. That's the default behaviour of the model, not something you opt into, and it's the single most confusing thing about running this pack for the first time - you ask for a prompt rewrite and get a blank reply with a wall of deliberation sitting in the Reasoning output. This node is the dial that controls how much of that thinking happens.

It takes three values: low, medium, xhigh. low is the default, and that's a deliberate choice by the pack's author rather than an afterthought: with a thinking model, a small max_tokens budget can be spent entirely in the hidden channel, so the lowest effort setting is what keeps the visible reply alive. If a Chat node returns nothing, the README says exactly what to do - raise Max Tokens or lower the effort. It's the same advice as the Max Tokens node, from the other end.

Inputs and output

• Reasoning Effort - one of low, medium, xhigh. (Yes, xhigh and not high; those are the strings the server takes, so don't go typing high into an Extra Body override.)
• Options - in and out, merged like every other node in the chain. Writes reasoning_effort.

What each setting is good for

low should be your default, and for the jobs this pack is built around it's the only setting you need. Rewriting a rough idea into a structured prompt, turning an image into a caption, reformatting text - none of that benefits from deliberation. The model doesn't need to reason about anything; it needs to follow a format and stop. Lower effort also means lower latency, and each token spent thinking is a token not spent on the answer, bought at the same price.

medium earns its keep when the task has an actual step in it - combine what's in these three reference images, or produce a shot list that has to stay internally consistent across entries.

xhigh is for problems where getting it wrong is expensive and you're willing to wait. The tooltip's warning is the honest one: it spends a lot of tokens on hidden reasoning. On a 27B model sharing a card with ComfyUI, that's minutes of wall clock, and with a 16k context it's a real slice of your budget gone before the answer starts.

This is one of the rare places where the community's usual line - reasoning models aren't the answer, the job wants small and obedient - gets a partial reprieve. The failure it warns about is a reasoner's scratch work leaking into your prompt as literal text, complete with "Here is your enhanced prompt:" and role delimiters. This pack sidesteps that structurally: the visible answer and the hidden reasoning arrive as two separate fields, so cranking the effort up can't contaminate the text you wire downstream. It can still eat your budget and your time, which is the reason to leave it on low anyway.

Where people get burned

Blank Response, non-empty Reasoning. The signature failure of this whole family. Check the node's preview for a reasoning-token count that ate your ceiling, then either set effort to low or raise Max Tokens.

Setting it on the wrong node. The chain merges settings, so if two Reasoning Effort nodes end up in one graph, the one whose output is later in the wiring is the one the Chat node sees. One node per graph; there's no reason for more.

Expecting the Reasoning output to be tidy. It's the model's raw working, not a formatted field. It's useful for debugging and occasionally for seeing why the model picked a certain reading of your prompt. Don't wire it into anything that expects prose.