Reasoning models and the R indicator

Short answer

Reasoning models (DeepSeek R1, Qwen QwQ, Qwen3-thinking) generate a hidden “thinking” block before their answer. Abliteration renders it as a collapsible card labeled “Thinking… (Ns)” during generation, “Thought for Xs” when done. Chats that use reasoning get a small R marker in the drawer.

Detailed explanation

What reasoning models do. Instead of jumping straight to an answer, these models first generate a chain-of-thought inside <think>...</think> tags. This can be hundreds or thousands of tokens of internal reasoning - exploring possibilities, checking work, discarding wrong paths. Only after </think> does the final answer begin.

How Abliteration renders it:

1. During generation - a card appears with header “Thinking… (Ns)” and a live counter of elapsed seconds. The reasoning text streams inside the card. The final answer streams below the card once </think> is emitted.

2. After generation - header changes to “Thought for Xs” (final elapsed time). The card is collapsed by default with a chevron toggle - click to expand and read the reasoning.

3. In the drawer - chats that have used reasoning show a small R marker to the left of the chat title. This helps you spot which chats are reasoning chats without opening them. If the same chat also used Council mode, the marker becomes Rᶜ - see Drawer navigation for how the composite reads.

Copy button. The reasoning card has its own copy icon in the header - copies only the reasoning text, separate from the final answer. Useful for citing what the model was thinking.

Which models trigger this. Any model that emits <think>...</think> tags in output. Common examples:

• DeepSeek R1 (all variants)
• Qwen QwQ 32B
• Qwen3-thinking variants
• Some fine-tunes explicitly trained for reasoning

Regular models don’t emit these tags - no card, just a plain answer.

Why it matters

Reasoning models often produce much better answers on hard problems (math, logic, code review) but at the cost of much longer generation times. Seeing the “Thinking…” card with elapsed counter lets you know the model is working, not stuck. The collapsible design keeps the chat clean - you don’t have to scroll through 2000 tokens of reasoning to find the final answer.

In Abliteration specifically

The reasoning card uses subtle grey borders - visually quiet, so it doesn’t dominate the chat.

Reasoning is saved to the database along with the message, so re-opening a chat later still shows the collapsed “Thought for Xs” card, not just the final answer.

If the same reasoning model is also used for chat title generation, the <think> blocks are stripped before storing - you never see them leak into titles. (Context compression summaries are ephemeral - never shown or stored - so they don’t matter for user-visible content.)

Related

• Reasoning persistence - how reasoning survives reload
• Drawer navigation - the R marker in chat list
• Reasoning not showing - when the card doesn’t appear