Qwen3 Technical Report
by Alibaba (Qwen Team)
The primary-source technical report behind the open-weight Qwen3 family, with its unified thinking / non-thinking design.
Overview
The Qwen3 Technical Report (arXiv:2505.09388, submitted May 14, 2025 by the Qwen Team at Alibaba) documents Qwen3, described in its abstract as 'the latest version of the Qwen model family... designed to advance performance, efficiency, and multilingual capabilities.' The series includes both dense and Mixture-of-Experts (MoE) architectures with parameter scales ranging from 0.6 to 235 billion. A defining contribution is the integration of a 'thinking' mode for complex, multi-step reasoning and a 'non-thinking' mode for rapid responses into a single unified framework, plus a 'thinking budget mechanism, allowing users to allocate computational resources adaptively during inference.' The report also details knowledge distillation from flagship models to smaller variants, expansion of multilingual support from 29 to 119 languages and dialects, and competitive results against models such as o1, o3-mini, and DeepSeek-V3 on coding, mathematics, and agent-related tasks. All models are released under Apache 2.0.
At a Glance
- Topic
- Models
- Level
- Advanced
- Format
- Paper
- Cost
- Free
- Duration
- ~1 hour read (technical report)
- Provider
- Alibaba (Qwen Team)
- Hands-on
- No
- Certificate
- None
What You’ll Learn
- ✓How Qwen3 unifies 'thinking' (multi-step reasoning) and 'non-thinking' (fast response) modes in one model
- ✓How the 'thinking budget' lets you trade inference compute against answer quality at runtime
- ✓The tradeoffs between dense and Mixture-of-Experts (MoE) architectures across 0.6B–235B parameters
- ✓How strong-to-weak knowledge distillation transfers capability to smaller, cheaper variants
- ✓How Qwen3 scales multilingual coverage to 119 languages and where it stands versus o1, o3-mini, and DeepSeek-V3
Highlights
- •Primary-source report from the Qwen Team, not a secondary summary
- •Covers a full open-weight lineup (Apache 2.0) from 0.6B to 235B, dense and MoE
- •Explains the thinking/non-thinking design and thinking-budget mechanism relevant to real serving and cost decisions
Who It’s For
Best For
- ✓AI engineers building on, fine-tuning, or serving open-weight LLMs
- ✓Practitioners choosing between dense and MoE models for reasoning workloads
- ✓Researchers and teams evaluating Qwen3 against o1/o3-mini/DeepSeek-V3
Prerequisites
- •Understanding of transformer LLM architectures and training
- •Familiarity with MoE, distillation, and inference-time reasoning
- •Comfort reading ML technical reports
FAQ
What is Qwen3 Technical Report?
The official arXiv technical report (arXiv:2505.09388) for Qwen3, Alibaba's open-weight LLM family spanning 0.6B–235B parameters. For AI engineers who build with, fine-tune, or serve open models and want the primary source on Qwen3's architecture, training, and reasoning modes.
Is Qwen3 Technical Report free?
Qwen3 Technical Report is free to access.
What level is Qwen3 Technical Report for?
Qwen3 Technical Report is aimed at a advanced audience. Recommended background: Understanding of transformer LLM architectures and training, Familiarity with MoE, distillation, and inference-time reasoning, Comfort reading ML technical reports.
How long does Qwen3 Technical Report take?
Expect roughly ~1 hour read (technical report). Most learners work through it at their own pace.
What will I learn from Qwen3 Technical Report?
You'll learn: How Qwen3 unifies 'thinking' (multi-step reasoning) and 'non-thinking' (fast response) modes in one model; How the 'thinking budget' lets you trade inference compute against answer quality at runtime; The tradeoffs between dense and Mixture-of-Experts (MoE) architectures across 0.6B–235B parameters; How strong-to-weak knowledge distillation transfers capability to smaller, cheaper variants; How Qwen3 scales multilingual coverage to 119 languages and where it stands versus o1, o3-mini, and DeepSeek-V3.