Reasoning model A model that thinks before it answers, and spends tokens doing it.
Short answer. A reasoning model is a language model trained to work through a problem step by step, in a long stretch of thinking text, before it gives its final answer. DeepSeek-R1 showed that reinforcement learning alone can build this habit: on the AIME 2024 maths contest, its R1-Zero model rose from 15.6% to 77.9% pass@1 during training. The price is time and tokens: the thinking is generated word by word, so the first visible word comes later and the thinking fills the context window. Reasoning helps with mathematics, code and multi-step analysis, and adds little to extraction, summaries or translation, where it is often switched off.
01
How it works
A reasoning model is trained, mostly by reinforcement learning, to produce a long working-out first and then a short answer. DeepSeek-R1’s rewards were rule-based: maths problems with a known final answer, and code that had to pass test cases, plus a reward for keeping the thinking inside think tags. The DeepSeek-R1 paper reports that the model learned to spend more thinking time as training went on, and to reflect on and check its own steps. In practice the output has two parts: a thinking block, often wrapped in think tags, and the answer.
Many models let you choose. Qwen3 has a thinking mode and a non-thinking mode, switched by the enable_thinking setting or by adding /think or /no_think to a prompt. OpenAI’s gpt-oss offers low, medium and high reasoning levels. Ollama’s think setting turns thinking on or off, or selects a level, and returns the thinking in a separate field from the answer.
02
Thinking versus non-thinking
Aspect
Non-thinking
Thinking
Output
The answer only
Thinking text, then the answer
First visible word
After the prompt is read
After the prompt is read and the thinking is written
Tokens used
Those of the answer
The answer plus all the thinking
Suits
Extraction, summaries, translation, formatting
Mathematics, code, planning, multi-step analysis
Qwen3 recommended settings
Temperature 0.7, top-p 0.8, top-k 20
Temperature 0.6, top-p 0.95, top-k 20; do not use greedy decoding
03
Costs and cautions
Time: thinking tokens are generated one by one at the model’s normal speed, so 3,000 thinking tokens at 10 tokens per second add five minutes before the answer starts.
Capacity: thinking text counts against the context window and uses KV cache, so fewer users fit at once.
Not proof: written reasoning does not always reflect why the model answered as it did. Treat it as a clue, and check the answer against sources.
Privacy: the thinking often repeats parts of the input. If you log it, handle the log with the same care as the documents themselves.
Settings: makers publish different recommended settings for thinking mode, and DeepSeek advises a temperature of 0.5 to 0.7 for its distilled models. Follow the model card.
04
Why it matters for an office
Decide per task, not per model. Pulling decisions and owners out of a meeting is extraction, which a non-thinking model handles quickly; ANVE’s published test for meeting minutes ran with thinking off. Checking a tender response against a long set of conditions, or reconciling figures across documents, is multi-step analysis where thinking can pay for its extra time. On a shared machine, thinking multiplies the tokens each request uses, so it also reduces how many requests the machine can serve at once.
05
Questions people ask
Are reasoning models more accurate?
On tasks that need several logical steps, such as maths and code, usually yes. On simple extraction or summaries they add time without much gain. Test on your own tasks.
Can I turn thinking off?
In many models, yes. Qwen3 has an enable_thinking setting and /no_think, and Ollama has a think parameter. Some reasoning models are built to always think.
Why is my reasoning model slow to start answering?
It is writing its thinking first. The answer begins only after the thinking ends, so the delay grows with how long the model thinks.
Can I see the reasoning?
Often yes. Ollama returns the thinking in its own field, and gpt-oss gives full access to its reasoning, although its model card says that is not intended to be shown to end users. Read it as a clue, not a guarantee.
06
What is not real yet
No sealed ANVE unit exists; every image of the appliance on this site is a labelled render. The software runs today on a development machine. No customers yet, no certifications, no delivery date, and the final price is not announced.
Reservations are ₹999 in India or US$12 elsewhere, refundable in full until you confirm an order in writing (terms). Offices that cannot use cloud AI can apply to be one of ten design partners.
