Ask a model “what is 17 × 24?” and it will answer instantly — sometimes correctly. Ask the same model with thinking mode on and it will work through the multiplication in steps first, then answer. The second answer is more likely to be right.
That is the whole idea, and it is one of the more genuinely useful developments in small models. It is also frequently misunderstood, misapplied, and left switched on when it should be off.
This article sits under How LLMs Actually Run on Your iPhone.
Why a model gets a problem wrong that it “knows” how to solve
A standard language model generates one token at a time, and each token takes a fixed amount of computation. There is no mechanism for spending longer on a hard question than an easy one. The word after “17 × 24 = ” gets exactly as much processing as the word after “the capital of France is ”.
That is a real structural limitation. Some problems genuinely require sequential steps, and a model forced to produce an answer immediately has to compress all of that into a single forward pass.
The fix turns out to be almost absurdly simple: let the model write its reasoning down first.
Once the model generates “First, 17 × 20 = 340,” that text becomes part of its own input for the next token. It can now build on it. The intermediate result is held in the text rather than having to be held implicitly inside one pass. Sequential computation becomes possible because the model gets more passes.
This was formalized in the chain-of-thought prompting work, which showed that simply prompting a model to reason step by step substantially improved multi-step accuracy. Reasoning models take the next step: they are trained to do it automatically, and trained on what good reasoning looks like — including backtracking and checking their own work.
What it looks like in practice
A reasoning model emits its reasoning inside a delimiter — commonly a <think> block — before the answer. Interfaces usually collapse or hide it.
For “a shop sells pens in packs of 12 and pencils in packs of 20; what is the smallest number of packs to get equal counts?”, a reasoning trace looks roughly like:
Need the least common multiple of 12 and 20.
12 = 2² × 3, 20 = 2² × 5.
LCM = 2² × 3 × 5 = 60.
60 ÷ 12 = 5 packs of pens.
60 ÷ 20 = 3 packs of pencils.
Check: 5 × 12 = 60, 3 × 20 = 60. Equal.Then the answer. The trace is not decoration — it is the computation. Remove it and the model has to do all of that in one shot, which is exactly where small models fail.
DeepSeek’s R1 distilled models made this visible-reasoning style widely available at small sizes, and the Qwen 3 and 3.5 families adopted a toggleable version of it.
What it costs
Reasoning tokens are ordinary tokens. They take the same time to generate as answer tokens, and they consume the same context.
If a model runs at 30 tokens per second and produces 900 reasoning tokens before a 150-token answer, that is 35 seconds — for an answer that would have taken 5 seconds without thinking. You waited seven times longer and saw the same amount of text.
There are three costs, and all of them are real:
Time. The dominant one. Reasoning can easily be 3–10× the answer length.
Context. Reasoning tokens occupy the context window like any other. Long reasoning across many turns fills a window considerably faster.
Battery. More tokens generated means more sustained computation, and on a phone that is measurable.
On a data centre this is cheap. On a device you are holding, it is a decision you actually feel — which is why a per-conversation toggle matters more on mobile than it does in a browser tab.
When to turn it on
Thinking mode earns its cost when a problem has multiple dependent steps and a checkable answer.
Worth it:
- Arithmetic beyond one operation
- Logic and constraint puzzles
- Debugging — tracing what a piece of code does line by line
- Planning with dependencies and ordering
- Comparing several options against several criteria
- Unit conversions and multi-stage calculations
- “Is this argument sound?” analysis
Not worth it:
- Factual recall. The model knows the capital of France or it doesn’t; reasoning cannot manufacture a fact.
- Summarizing and rewriting. The work is in reading, not deducing.
- Creative writing. Reasoning tends to make prose more stilted, not better.
- Translation.
- Casual conversation.
- Anything where you want an answer in under five seconds.
A reasonable default on a phone is off, switched on deliberately when a question is genuinely hard. Most questions aren’t, and the latency is not free.
Why this matters more for small models than large ones
Here is the part that makes thinking mode particularly interesting on a phone.
Extra reasoning helps small models proportionally more than large ones. A large model has enough capacity to handle several implicit steps within a single pass; a 2B model does not. Giving that 2B model explicit steps to work through recovers a meaningful part of the gap.
It does not close it. A 2B model with thinking will not match a frontier cloud model on genuinely hard reasoning, and claiming otherwise would be nonsense. But on the middle band of problems — the multi-step-but-not-research-grade questions that make up most practical use — thinking mode moves a small local model from “unreliable” to “dependable.”
That is a large practical difference. It means the accuracy trade-off for choosing a local model is smaller than the parameter counts suggest, which connects directly to the argument in Small Language Models.
Using it well in Cloaked
The Qwen 3 and Qwen 3.5 families and DeepSeek R1 1.5B all support thinking in Cloaked, and the Qwen models let you toggle it per conversation.
A practical pattern:
- Keep thinking off for a general-purpose model you use all day.
- Switch it on when you hit something that needs working through.
- Read the reasoning when the answer surprises you — it usually shows exactly where things went wrong, which is more useful than a wrong answer with no explanation.
That last point is underrated. A visible reasoning trace makes a model’s mistakes legible. A cloud assistant that confidently states a wrong figure gives you nothing to check. A local model that shows “60 ÷ 12 = 6” lets you spot the error immediately.
For how reasoning models compare to the rest of the catalogue, see Best Local LLM Models for iPhone.
Download Cloaked on the App Store to try thinking mode on your own device — the reasoning happens on your phone, and neither the reasoning nor the answer is ever transmitted.
Frequently asked questions
What is thinking mode in an AI model?
It is a mode where the model generates its reasoning as tokens before producing a final answer. Those intermediate tokens are usually hidden or collapsed in the interface. The reasoning becomes part of the model's own input, letting it build on earlier steps rather than answering in one pass.
Does thinking mode make answers better?
On multi-step problems, clearly yes — arithmetic chains, logic puzzles, planning, and debugging all improve. On simple factual questions, summarization, or creative writing it adds little and sometimes hurts, since the model may overthink something that needed a direct answer.
Why does thinking mode take so long?
Reasoning tokens are ordinary generated tokens and cost the same time each. A model producing 800 tokens of reasoning before a 200-token answer takes five times as long as answering directly, even though you only see the final part.
Which models in Cloaked support thinking mode?
The Qwen 3 and Qwen 3.5 families and DeepSeek R1 1.5B all support a thinking toggle. On the Qwen models it can be switched on and off per conversation, so you can reserve it for problems that benefit.