# Choosing when a model thinks longer.

The Qwen team released and described Qwen3, a family of dense and mixture-of-experts language models with selectable thinking and non-thinking modes.

Canonical: https://brightaifuture.com/discoveries/qwen3
Format: discovery
Source publication: 2025-04-29
Bright publication: 2026-09-07
Substantive update: None recorded
Evidence and review: Deployed; confidence: unassessed; approved; ai-assisted. AI-assisted editorial comparison with the cited primary source; result, setting, source date and limitations retained. Independently checked within the research team. Publication authorized by the site owner; no human source review is claimed.

## The human problem

People building AI systems often need to balance response speed and cost against extra time for complex reasoning, including in multilingual work.

## The prior constraint

A system's reasoning budget is often fixed or hidden from the person using it.

## AI’s actual role

Qwen3 can use a slower step-by-step Thinking Mode or a faster Non-Thinking Mode; the team also reports support for 119 languages and dialects.

## The documented result

Qwen announced models from 0.6B to 32B dense sizes and 30B-A3B/235B-A22B mixture-of-experts sizes, with pre- and post-trained variants available via Hugging Face, ModelScope, and Kaggle. Its source describes the two operating modes and reports multilingual support.

## Why it may matter

Giving developers a stated control over reasoning time could make it easier to reserve slower processing for harder tasks and use faster responses for straightforward work.

## Limitations

The source's model rankings and language-support claims do not establish comparable accuracy or cultural fit across every listed language or task.

The source says the family was open-sourced but this dossier does not assert a single license or complete training-data release for every variant; those must be checked per model card.

Longer reasoning can increase latency and cost without guaranteeing a correct answer.

## Unresolved questions

Do native speakers judge quality and safety similarly across the claimed languages?

When does additional reasoning actually improve a real task?

What data, licensing, and provenance constraints apply to each released variant?

## Provenance and history

{
  "dates": {
    "eventDate": null,
    "publicationDate": "2025-04-29",
    "captureDate": "2026-09-07",
    "lastReviewedDate": "2026-09-07"
  },
  "provenance": {
    "origin": "editorial",
    "externalId": "https://qwenlm.github.io/blog/qwen3/"
  },
  "revisions": [
    {
      "id": "revision:b4305dfe89bc751fb480",
      "recordedAt": "2026-09-07",
      "summary": "Giving developers a stated control over reasoning time could make it easier to reserve slower processing for harder tasks and use faster responses for straightforward work.",
      "sourceIds": [
        "source-qwen3"
      ]
    }
  ],
  "corrections": []
}

## Original sources

- [Qwen3: Think Deeper, Act Faster](https://qwenlm.github.io/blog/qwen3/)

## Continue exploring

- [Open intelligence](https://brightaifuture.com/worlds/open)
- [What changes when powerful models become open-weight?](https://brightaifuture.com/threads/open)
- [Someone builds on it](https://brightaifuture.com/open-intelligence)