What is Qwen3, and what does "hybrid thinking" mean?

Alibaba's Qwen team has released Qwen3, the newest generation of its large language models. The announcement went out on Monday evening, April 28, in the United States, which was already Tuesday, April 29, in China, where the blog post is dated. The weights, the trained numbers that make up each model, are free to download under the Apache 2.0 license, which allows commercial use.

The family covers almost every size you could want, from a model of under a billion parameters up to a flagship that Alibaba compares with Google's Gemini 2.5 Pro, xAI's Grok 3, DeepSeek-R1 and OpenAI's o1 and o3-mini. Its signature feature is what the team calls hybrid thinking: one model that can either reason at length or answer at once.

What Alibaba released

There are two kinds of models in the family. Six are dense models, where every part of the network works on every word. Two are mixture-of-experts (MoE) models, explained below.

Model	Type	Total parameters	Active per word	Context length
Qwen3-235B-A22B	MoE	235 billion	22 billion	128K tokens
Qwen3-30B-A3B	MoE	30 billion	3 billion	128K tokens
Qwen3-32B, 14B, 8B	Dense	32, 14, 8 billion	all	128K tokens
Qwen3-4B, 1.7B, 0.6B	Dense	4, 1.7, 0.6 billion	all	32K tokens

A parameter is one of the adjustable numbers a model learns during training; more parameters usually means more capability and more memory needed. The context length is how much text the model can consider at once, measured in tokens, pieces of words. The names tell you the sizes: "235B-A22B" means 235 billion parameters in total, 22 billion active at a time.

The models are on Hugging Face, ModelScope and Kaggle, and you can chat with them on the Qwen Chat website and app. TechCrunch notes that cloud providers such as Fireworks AI and Hyperbolic already offer them as well.

How hybrid thinking works

Reasoning models, such as OpenAI's o-series and DeepSeek-R1, write out a long chain of intermediate steps before they answer. That helps with math, code and logic, but it is slow and uses more computing for simple questions. Until now, you usually chose a different model for each job.

Qwen3 puts both behaviors in one model. In thinking mode it works through the problem step by step; in non-thinking mode it replies right away. You can switch per message by adding /think or /no_think to your prompt, and the model follows the latest instruction. Developers can also turn thinking on or off in code and set a thinking budget, a cap on how long the model reasons. The Qwen team reports that quality rises smoothly as that budget grows, so you can trade speed for care task by task.

The team built this in four training stages. First it taught the model long chains of reasoning with examples. Then it used reinforcement learning, rewarding answers that check out, to strengthen that reasoning. Third, it blended in ordinary quick-answer training. Finally, it ran reinforcement learning again on more than 20 everyday tasks, such as following instructions and formats.

Mixture of experts is the other idea worth knowing. Instead of one huge network, an MoE model has many smaller sub-networks called experts (128 in Qwen3's MoE models) and a router that sends each word to just a few of them (8 here). Picture a large hospital: every patient sees only the specialists they need, not the whole staff. That is why the 30-billion-parameter model runs with only 3 billion parameters active, and why it can be fast.

A worked example: switching modes

Here is a short conversation in Qwen Chat or any app running Qwen3:

Prompt

A train leaves at 9:40 and the trip takes 2 hours 35 minutes,
including a 12-minute stop. When does it arrive? /think

The model shows its working, adds the times, and answers 12:15. Then:

Prompt

Thanks. Now translate "the train arrives at a quarter past twelve"
into French and Spanish. /no_think

It translates immediately, with no reasoning step, because the task is simple.

Try it: Ask Qwen3 the same tricky question twice, once ending with /think and once with /no_think. Compare the answers and how long each took. That shows you when the extra thinking is worth the wait.

Why it matters

Qwen3 is notable for three reasons. First, its range: the same design and the same switch work from 0.6 to 235 billion parameters, so a developer can test on a laptop and scale up later. The team says even the small models punch above their weight; it reports Qwen3-4B rivals its previous 72-billion-parameter model.

Second, its reach. Qwen3 supports 119 languages and dialects and was trained on about 36 trillion tokens, roughly twice the data of its predecessor, Qwen2.5. The team also says it improved agent skills and support for MCP, the Model Context Protocol that lets models connect to outside tools.

Third, it adds to a trend. TechCrunch judges that no Qwen3 model clearly beats OpenAI's newest o3 and o4-mini, but calls them strong performers, and an industry executive it quotes sees Qwen3 as more evidence that open models are keeping pace with closed ones. All benchmark claims so far come from Alibaba itself; independent tests are likely to follow in the coming weeks.

How can you run Qwen3 on your own computer?

The easiest route is a free app that runs models locally, such as Ollama or LM Studio, both recommended by the Qwen team. With Ollama installed, one command such as ollama run qwen3:30b-a3b downloads the model and opens a chat in your terminal. LM Studio offers the same with a point-and-click window.

Pick a size your machine can handle. The smallest models (0.6B, 1.7B and 4B) need the least memory; the 8B and 14B models need more; the 30B MoE model is a good balance on a well-equipped computer because so few of its parameters are active at once. The flagship 235B model needs server-class hardware, so most people will try it on Qwen Chat or through a cloud provider instead.

Once it runs, everything stays on your machine, which is useful for private documents. Start with short tasks, use /no_think for quick ones, and check facts as you would with any model.