Alibaba releases Qwen3 with hybrid thinking modes

Background

Qwen2.5 was the previous generation of the Qwen line. Its pre-training used about 18 trillion tokens, and Qwen3 was trained on roughly double that amount.^[1]

What happened

On 29 April 2025 Alibaba released Qwen3 under the Apache 2.0 license.^[1][2] The family contains two mixture-of-experts models, Qwen3-235B-A22B and Qwen3-30B-A3B, and six dense models at 32B, 14B, 8B, 4B, 1.7B and 0.6B.^[1]

Key details

The defining feature is hybrid thinking. A single model can reason step by step or answer instantly. Users switch the mode with enable_thinking or with /think and /no_think in their prompts.^[1]

Pre-training covered about 36 trillion tokens and 119 languages and dialects.^[1]

The Qwen team said Qwen3-235B-A22B was competitive with DeepSeek-R1, o1, o3-mini, Grok-3 and Gemini-2.5-Pro on coding, math and general benchmarks.^[1] The release also strengthened agent and coding abilities and improved support for the Model Context Protocol (MCP).^[1]

Some terms are worth explaining. A dense model uses all of its parameters for every token. A mixture-of-experts model activates only part of its parameters, which is why the 235B model has 22 billion active parameters. Hybrid thinking refers to a single model that can produce a step-by-step reasoning process before its answer, or skip that process when a quick answer is enough.^[1]

Model Context Protocol (MCP) is a standard that lets a model connect to outside tools and data sources in a consistent way. The support mentioned in the release means that Qwen3 can be used inside software that speaks this protocol.^[1]

Pre-training is the stage in which a model learns from a large text corpus. The 36 trillion tokens used here were far more than the 18 trillion used for Qwen2.5, which is the comparison the team drew.^[1]

Impact

Qwen3 brought the thinking and non-thinking modes into one model family. That design was changed in the next update: in July 2025 Qwen3-2507 split the two modes into separate instruct and thinking models.^[1]