Reasoning mode
See the AI think before it answers. Toggle on for hard scenes, off for casual chat.
Several Reverie models can show their reasoning — the private train of thought they go through before composing a reply. Reasoning makes characters more deliberate, more in-character, and (sometimes) much slower. You decide per model whether you want it on.
Three flavours of reasoning support
| Capability | Models | What it means |
|---|---|---|
| None | DeepSeek V3.2, Gemini 3 Flash, Llama 3.1 8B, Qwen3 VL | No reasoning support. Standard reply only. |
| Optional | MiMo V2 Flash, DeepSeek V4 Flash, GLM 4.7, GLM 5, GLM 5.1 | You toggle reasoning on or off per chat. |
| Always | DeepSeek R1 | Always reasons. Slower replies, denser thinking. |
Turning reasoning on
Open the model picker at the top of the chat. Each "optional reasoning" model shows a small 🧠 Reasoning toggle. Tap to enable.
When reasoning is on, you'll see a collapsed pane labelled Thinking… appear with each AI reply. Tap to expand and read the chain of thought.
Want reasoning always expanded for the latest reply? Settings → Chat advanced → Auto-expand reasoning for last message.
When to turn it on
| Scenario | Reasoning helps? |
|---|---|
| Casual chat, daily catch-up | No — slower for no benefit |
| A character solving a puzzle in-fiction | Yes — they think it through visibly |
| Emotional confrontation that needs subtext | Yes — the model considers what they would not say |
| Tutoring / explanation | Yes — the model checks itself |
| Hot-take humour, banter, fast pacing | No — reasoning slows the joke |
What it costs
Reasoning tokens count toward credit consumption just like reply tokens. A reasoned reply typically costs 1.5-2× a non-reasoned reply on the same model. The DeepSeek R1 (always-reasoning) model is priced at a 1.0× multiplier already factoring that in.
What it doesn't do
Reasoning is not memory. The thoughts visible in the reasoning pane are scratch — they don't persist between turns or affect long-term memory. If you want something remembered, pin it.
Reading the reasoning pane
The text inside isn't dialogue — it's the model talking to itself. It will often say things like:
The user just shared their dad died. Don't jump to platitudes. Pull on the earlier conversation about how they said they're not religious. What would Eliza say? She would say nothing for a moment.
Reading it can be illuminating — and occasionally jarring. If you'd rather not see it, leave the toggle off. The reply itself doesn't change based on whether you read the reasoning.