What is a reasoning model?
Some models answer straight away. Others generate a long stretch of working first. Here is what that thinking is, and when it is worth waiting for.
- → A reasoning model writes out intermediate steps before it commits to an answer.
- → The steps are still predicted text, not a separate kind of logic, but they measurably improve hard multi-step problems.
- → Thinking costs time and tokens, so it pays off on hard questions and wastes both on easy ones.
Ask a standard chat model a question and it starts answering immediately, one token after another. A reasoning model does something different: it first produces a stretch of working, often long, where it breaks the problem down, tries approaches, and checks itself. Only then does it write the answer you see.
That working is usually called thinking or chain of thought. DeepSeek, Kimi, GLM, and Qwen all do it, as do the closed frontier models. It is the biggest change in how language models work since chat itself.
Thinking is still text prediction
Nothing new happens inside the model when it thinks. It is predicting the next token exactly as before. The difference is that it was trained to spend those tokens on scratch work: restate the problem, list what is known, try a path, notice a contradiction, back up. Because every later token can see everything written so far, that scratch work acts as working memory the model would not otherwise have.
Why writing steps out helps
A plain model has to get from question to answer in one pass through its layers. A hard problem with several dependent steps does not fit in one pass. Writing the steps out turns one impossible leap into many small, easy ones. That is why reasoning models jump ahead on math, code, logic puzzles, and planning, and barely move on tasks like casual writing that were never multi-step in the first place.
How they are trained to do it
Early on, people got some of this effect just by asking a model to think step by step. Reasoning models take it further: during training they are rewarded when a long chain of working ends in a checkable correct answer, so the habit of careful working gets reinforced. The result is a model that thinks by default, and that has learned which kinds of working actually lead somewhere.
What it costs
Thinking is slow and not free. A reasoning model may generate thousands of tokens of working before its first visible word, which means a longer wait and a higher bill per question. On a simple factual question or a quick rewrite, that is pure waste: the answer is no better and it arrives later. The skill is knowing which questions deserve it.
Reading the thinking
Some products show the working, some show a summary, some hide it. Reading it can be useful: you can spot a wrong assumption early. But treat it as a rough trace, not a faithful report of what the model really did. The visible steps can look tidy while the answer actually came from somewhere else.
It is the difference between answering an arithmetic question in your head and being handed a sheet of paper. With paper you can write down partial results, check them, and get harder sums right. The paper does not make you smarter. It lets you use what you know without losing your place.
Every model on Berges AI is a reasoning model, and you watch the working stream before the answer lands. You pick which one from the sidebar: DeepSeek is the quickest, and GLM, Qwen and Kimi are the big ones for the hardest questions.
Try Berges AIRelated concepts
Things people ask
Is a reasoning model the same as an agent?
No. A reasoning model thinks before it answers but still just produces text. An agent is a model wired up to tools and a loop so it can act: search, run code, call a service. Many agents use a reasoning model underneath, but the two ideas are separate.
Does thinking stop hallucinations?
It reduces the errors that come from skipping steps, which is a lot of them on technical questions. It does nothing for a fact the model never learned. A reasoning model can work carefully from a wrong premise and land on a confident wrong answer, so check anything that matters.
Should I always use the reasoning model?
No. For a quick question or a short piece of writing it is slower, costlier, and not better. Use it when the problem has several steps that depend on each other: math, debugging, a decision with trade-offs, anything you would want to see worked out.
Can I switch reasoning off on Berges AI?
No. Every model on Berges AI reasons before it answers. On an easy question the working is short, and on a hard question the working is exactly what you want to see.