Qwen3

Thinking costs tokens

When a model writes out reasoning, each reasoning token becomes context for the tokens after it. Writing intermediate steps therefore gives the model more computation before it commits to an answer. The steps can also contain mistakes, repeat themselves or build on a false premise. The length of the reasoning measures how much was spent, not whether the answer is right.

The original Qwen3 models can answer in a thinking mode or a non-thinking mode, and the announcement describes an adjustable budget for how much thinking to allow. The family includes six dense models, from 0.6 to 32 billion parameters, and two mixture-of-experts models. In a name such as 235B-A22B, 235 billion is the total number of parameters and 22 billion is roughly the number active for each token. [1]

Take 17 × 23. A direct reply can just output 391. A thinking reply can split 23 into 20 + 3, compute 340 + 51, and then answer 391. A bare answer does not reveal how it was produced. Writing out the steps makes the intermediate results available to later tokens, but it does not make the final answer infallible.

Four training stages, then distillation

The Qwen3 report trains its flagship models in four post-training stages. First, the model learns from examples of long reasoning. Second, reinforcement learning on problems with checkable answers strengthens that reasoning. Third, thinking-mode fusion trains on a mix of thinking and non-thinking examples so that one model can do both on request. Fourth, general reinforcement learning improves behavior across everyday tasks. The smaller models skip most of this. They learn from the large models by strong-to-weak distillation, which the report says took about a tenth of the GPU-hours of running all four stages. [2]

Each stage teaches something different. Examples show what a successful reasoning trace looks like. A checker judges whether the model’s own attempt reached a correct result. Fusion teaches that the same kind of question may call for a short answer or a long deliberation, depending on the request. General training covers tasks that have no single correct answer.

A student learning algebra needs the same mix: a worked example to show the format, new problems with an answer key for feedback, and a separate lesson on when a one-line answer is enough. More of one kind of lesson does not replace the others.

What a teacher’s probabilities add

A training label that says only “the next token is A” is a hard target. A teacher model’s probability distribution says more: A is likely, B somewhat likely, C and D unlikely. Qwen3’s distillation has two phases. First, each small model is trained on responses the large teacher wrote. Then the small model writes its own responses, and at each token it is trained to match the teacher’s probabilities for that position, computed from the teacher’s logits, the raw scores that become probabilities. Because this second phase uses the student’s own responses, the student is corrected in the situations it actually reaches, including after its own mistakes. [2]

The figure uses made-up scores of 4, 2, 1 and 0 for four tokens. The softmax function turns a score z into a probability proportional to exp(z / T), where T is the temperature. Raising T spreads probability across the alternatives; lowering it concentrates probability on the top choice. The slider changes this example only, not Qwen3’s training settings.

Suppose the student puts almost all its probability on B. A hard target says the answer is A, but says nothing about how B, C and D compare. The teacher’s distribution conveys that too. A teacher that is confidently wrong passes on its error just as faithfully, so distillation transfers behavior without being an independent source of truth.

Switching modes in practice

The Qwen3-8B model card shows two controls. A setting in the chat template, enable_thinking, turns thinking on or off for a request. When thinking is on, a user can also write /think or /no_think in a message to switch modes turn by turn. The template wraps the reasoning in special tokens so that an application can hide or show it. These controls describe the original April release; later models with Qwen3 in their names do not all work the same way. [3]

This matters when comparing results. If one benchmark run turns thinking off and another allows long reasoning, the two scores reflect different amounts of computation. The report enforces a thinking budget by stopping the reasoning when the budget runs out, inserting an instruction to answer now, and letting the model write its answer. Cutting off the whole response at the same length could instead stop before any answer appears. [2]

A useful evaluation records whether the final answer was correct, how many tokens were generated, how long it took and whether an answer arrived at all, on both easy and hard questions. The question is whether the extra computation improves results enough to be worth its cost, and whether a smaller distilled model keeps that balance.

Simplifications and limits

• The written reasoning is not a complete or faithful account of the model’s internal computation.
• The softmax figure shows what information distillation carries. It does not train anything or predict a student’s quality.

Follow the idea

Qwen2.5Qwen3-Next

Sources and further reading

Primary papers, model cards, and release posts. Facts describe the versions named here; the interactive calculations are teaching models. Research cutoff: October 3, 2026.

1. Qwen3: Think Deeper, Act FasterPublic release, hybrid thinking, model sizes and Apache 2.0 weights.
2. Qwen3 Technical ReportSections 4.1–4.5 and 4.7; four-stage post-training and on-policy teacher-logit distillation.
3. Qwen3-8B model cardOriginal model chat-template controls and thinking-mode usage.