Paper summary: Qwen3 Technical Report

Problem

Users had to pick between chat models (fast) and dedicated reasoning models (slow but careful), often from different model families.

Method

Qwen3 is a family of dense and MoE models from 0.6B to 235B parameters with:

• thinking mode (multi-step reasoning) and non-thinking mode (quick answers) in one model, switchable per query or through the chat template
• a thinking budget, so users can choose how much compute to spend at inference
• smaller models built with knowledge from the flagship models, which cuts their training cost

Key results

From the abstract: state-of-the-art results across benchmarks including code generation, mathematical reasoning and agent tasks, competitive with larger MoE and proprietary models. Multilingual support grows from 29 to 119 languages and dialects compared with Qwen2.5. All models are released under Apache 2.0.

Limitations

There’s no dedicated limitations section in the extracted text. Results are reported by the developers.