Reasoning Effort

Reasoning effort controls how hard the model thinks before it writes a final answer. Character Vault sends this as part of the chat request when Enable reasoning is on (Settings → AI Config → Advanced Options).

This page explains what each level means, which models use which levels, and how Character Vault handles providers that only accept a subset of values.

Enable reasoning first

Setting	What it does
Enable reasoning	Turns on thinking / chain-of-thought for models that support it. Off by default.
Show reasoning	When on, the model’s thinking is shown in a collapsible section before the answer.
Reasoning Effort	How much thinking budget to request (only shown when reasoning is enabled).

Reasoning is useful for harder tasks: multi-step rewrites, consistency checks, complex character logic, or long instructions. It is usually slower and uses more tokens (and cost) than a normal reply.

TIP

If you only need quick grammar or short polish, leave reasoning off. Turn it on when quality matters more than speed.

What the levels mean

Character Vault offers six effort values. Higher effort generally means more thinking tokens, longer waits, and often better answers on hard problems. Exact behavior is decided by the model and the API gateway (NanoGPT, OpenRouter, OpenAI, etc.).

Level	In short	Typical use
Minimal	Almost no thinking. Fastest.	Tiny edits, classification-style tasks, when you want speed.
Low	Light thinking. Still quick.	Simple rewrites, short expansions.
Medium	Balanced depth and speed. Default.	General toolbar and chat work.
High	Thorough thinking.	Complex edits, long context, careful reasoning.
Extra high (xhigh)	Peak depth on OpenAI-style effort ladders.	Hardest tasks on GPT-5.x / o-series via OpenAI or OpenRouter.
Max	Peak depth on many open / SOTA APIs.	Hardest tasks on DeepSeek V4, GLM-5.x, Kimi K3, and similar.

Why both “Extra high” and “Max”?

Providers do not share one enum:

• OpenAI-style APIs often use minimal → low → medium → high → xhigh (and sometimes none).
• DeepSeek V4, GLM-5 / 5.2, and several other SOTA stacks use a shorter ladder, often only high and max (plus turning thinking off).
• Kimi K2.5 / K2.6 often only toggle thinking on or off; multi-level effort may be ignored or rejected. Kimi K3 may only accept max today.

Character Vault exposes both Extra high and Max so you can pick the name your model’s API expects. You do not need to memorize every provider.

Which models use which levels?

This is a practical guide, not a guarantee. Gateways can remap or restrict values.

GPT and “OSS-style” ladders

Models that follow an OpenAI-like effort scale (many GPT-5.x, o-series, and some open weights served with the same API shape) usually understand:

• Minimal, Low, Medium, High for everyday control
• Extra high (xhigh) for maximum depth on that ladder

Medium is a safe default for these models.

SOTA / thinking-first models (DeepSeek, GLM, Kimi, etc.)

Newer flagship thinking models often do not support the full low/medium ladder:

Family (examples)	Typical effort values	Notes
DeepSeek V4 (thinking / :thinking variants)	high, max (and sometimes none to disable)	Official APIs often map low/medium → high and xhigh → max. Some gateways instead reject medium and list only none, high, max.
GLM-5 / GLM-5.2	high, max (max is often the model default)	Use High for cost/latency, Max for hardest agentic or coding-style work.
Kimi K2.5 / K2.6	Thinking on/off more than multi-level effort	Effort dropdown may have little effect; enable reasoning still matters for thinking output.
Kimi K3	Often max only today	Pick Max when using these IDs.
OpenRouter unified reasoning.effort	Can include none, minimal, low, medium, high, xhigh, max	Per-model support varies; OpenRouter may map your choice to what the upstream model allows.

Quick rule of thumb

You are using…	Prefer…
GPT / o-series / OpenAI-style	Minimal → High, or Extra high for max depth
DeepSeek V4, GLM-5.x, many SOTA open models	High or Max
Not sure	Start with Medium or High; if the request fails on effort, Character Vault remaps for you

Automatic remapping

Character Vault always sends the effort you selected when reasoning is on. If the provider returns an error like “unsupported reasoning_effort” and lists allowed values, Character Vault:

1. Picks the closest supported level (never maps down to “no thinking” while reasoning is enabled).
2. Treats Max and Extra high as peers when only one of them is allowed.
3. Remembers rejections for that base URL + model for the rest of the session so the next request does not fail the same way.

You may see a console message such as remapped parameters: reasoning_effort=high. That means your setting was adjusted for compatibility, not that reasoning was turned off.

If a parameter is completely unsupported (not just a bad value), Character Vault may drop it and retry. Non-standard sampler fields (min_p, top_k, repetition_penalty) can be dropped the same way on strict OpenAI-compatible hosts; that is separate from effort.

Practical recommendations

Goal	Suggestion
Everyday character editing with a GPT-style model	Medium
Faster toolbar passes	Low or Minimal, or turn reasoning off
Hard lore / multi-field consistency	High
DeepSeek / GLM hard reasoning	High first; Max if answers are shallow
OpenAI peak reasoning	Extra high when the model supports it
Cost / latency sensitive	Lower effort, or disable reasoning

Higher effort does not always improve creative prose. For roleplay flavor and style, sampler settings (temperature, top-p, etc.) often matter more. Use high effort when the task is logical or multi-step, not only when you want “more words.”

Related settings

• AI Setup — base URL, API key, model, streaming, NanoGPT / Synthetic / OpenRouter usage cards
• Sampler Settings — temperature, top-p, min-p, context length
• Editor & AI Toolkit — toolbar operations that use these settings
• AI Assistant Orion
• AI Agent

Finding this in the app

1. Open a character workspace.
2. Settings → AI Config → Advanced Options.
3. Turn on Enable reasoning.
4. Choose Reasoning Effort.
5. Click Save Settings.