Qwen3 Technical Report
by Qwen Team, Alibaba Cloud
The primary source on unified thinking/non-thinking modes and the thinking-budget knob.
Overview
The Qwen3 Technical Report (arXiv:2505.09388) is the Qwen team's own account of the Qwen3 model family, submitted on 14 May 2025 with 60 listed authors from Alibaba Cloud, and it is the primary source for how the series is built rather than a launch blog. Qwen3 spans both dense and mixture-of-experts architectures from 0.6 to 235 billion parameters: dense at 0.6B, 1.7B, 4B, 8B, 14B and 32B, and MoE as Qwen3-30B-A3B and the flagship Qwen3-235B-A22B, where the A-suffix gives activated rather than total parameters. The report's headline contribution is unifying a thinking mode for complex multi-step reasoning and a non-thinking mode for rapid context-driven responses inside a single framework, which removes the usual split between a chat-optimized model and a separate dedicated reasoning model; mode switching is driven by the user query or the chat template. It also introduces a thinking budget, a mechanism that lets a caller allocate reasoning compute adaptively per request and so trade latency against accuracy by task complexity. A third thread is using the flagship models' knowledge to build the smaller ones, cutting the compute needed to train them while keeping them competitive. Multilingual support expands from 29 languages and dialects in Qwen2.5 to 119. All Qwen3 models are released publicly under Apache 2.0, and the matching code and checkpoints live at github.com/QwenLM/Qwen3.
At a Glance
- Topic
- Models
- Level
- Advanced
- Format
- Paper
- Cost
- Free
- Duration
- ~2 hours to read in full; ~15 min for the architecture and mode-switching sections
- Provider
- Qwen Team, Alibaba Cloud
- Hands-on
- No
- Certificate
- None
What You’ll Learn
- ✓How thinking and non-thinking modes are unified inside a single model checkpoint
- ✓What a thinking budget is and how it trades inference latency against answer accuracy
- ✓The dense and MoE variants in the family and what activated parameters actually mean
- ✓How knowledge from flagship models cuts the compute cost of training smaller ones
- ✓Why unifying reasoning and chat removes a model-routing decision from your stack
- ✓The scope of Qwen3's multilingual expansion from 29 to 119 languages and dialects
- ✓How an Apache-2.0 open-weight family is positioned against larger proprietary models
Highlights
- •Primary source written by the model authors, not a secondary write-up or benchmark roundup
- •Documents the thinking-budget mechanism, which few open-weight model reports describe at all
- •Covers the entire 0.6B to 235B family in one place, dense and mixture-of-experts together
- •Everything described ships under Apache 2.0, so the claims are independently reproducible
- •Backed by a public repo (QwenLM/Qwen3, 27k+ stars) carrying the checkpoints and citation
Who It’s For
Best For
- ✓Engineers evaluating open-weight models for self-hosted or on-premises deployment
- ✓Researchers studying reasoning-mode unification and inference-time compute control
- ✓Teams sizing a mixture-of-experts versus dense tradeoff against their own hardware
Prerequisites
- •Comfortable reading ML research papers and interpreting benchmark tables
- •Familiarity with transformer training, MoE routing and post-training terminology
FAQ
What is Qwen3 Technical Report?
The Qwen team's own technical report on the Qwen3 family, covering dense and mixture-of-experts models from 0.6B to 235B parameters, the unification of reasoning and chat modes into a single checkpoint, and the thinking-budget mechanism. Afterwards you can reason about open-weight model selection, MoE sizing and inference-time compute control from the source rather than from benchmark blog posts.
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: Comfortable reading ML research papers and interpreting benchmark tables, Familiarity with transformer training, MoE routing and post-training terminology.
How long does Qwen3 Technical Report take?
Expect roughly ~2 hours to read in full; ~15 min for the architecture and mode-switching sections. Most learners work through it at their own pace.
What will I learn from Qwen3 Technical Report?
You'll learn: How thinking and non-thinking modes are unified inside a single model checkpoint; What a thinking budget is and how it trades inference latency against answer accuracy; The dense and MoE variants in the family and what activated parameters actually mean; How knowledge from flagship models cuts the compute cost of training smaller ones; Why unifying reasoning and chat removes a model-routing decision from your stack; The scope of Qwen3's multilingual expansion from 29 to 119 languages and dialects; How an Apache-2.0 open-weight family is positioned against larger proprietary models.
Topics
Sources
This page was written from 3 sources, 2 on domains other than arxiv.org.