How do you try Qwen3 and switch its thinking on and off?

Qwen3 is the newest family of open-weight models from Alibaba's Qwen team, released at the end of April 2025 under the Apache 2.0 license, which allows commercial use. "Open-weight" means the trained model files are published, so you can download and run them yourself. Its most unusual feature is that each model has two modes in one. In thinking mode it reasons step by step before answering, which helps with math, logic and code. In non-thinking mode it answers directly and quickly, which suits everyday questions.

You choose the mode, and you can change it from one message to the next.

Three ways to try it

1. In the browser. The Qwen team's own chat site, chat.qwen.ai, needs no installation.
2. With Ollama, a command-line tool for running models locally. The Qwen3 announcement gives this command for the 30B-A3B model: ollama run qwen3:30b-a3b.
3. With LM Studio, a desktop app. Search for "qwen3" in its Discover tab and pick a size that fits your memory (see the table below).

The model card also names MLX-LM (for Apple Silicon Macs), llama.cpp and KTransformers as supported local tools.

Worked example: the two switches

Thinking is on by default. The soft switch lets you change it per message: add /think or /no_think at the end of a prompt or in the system message, and the model follows the most recent instruction in the conversation.

Try this three-message sequence in any of the tools above:

A shop sells pens at 3 for $2. How much do 10 pens cost if you must buy them in packs of 3? /think

You should see a block of reasoning, wrapped in <think> tags or shown in a separate panel depending on the app, followed by the answer: four packs, $8.

Write a one-line slogan for that shop. /no_think

The reply comes back almost immediately. The model still outputs a thinking block, but it is empty.

Is buying 4 packs really the cheapest way to get at least 10 pens? /think

Reasoning returns for the check. Notice the pattern: switch thinking on for problems with several steps or a right answer to verify, and off for writing, chat and simple lookups, where it only adds waiting time.

The hard switch, in code

If you run Qwen3 with Hugging Face's transformers library (version 4.51.0 or later), you can turn thinking off completely when building the prompt:

text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
    enable_thinking=False,  # no reasoning block at all
)

With enable_thinking=False, the /think and /no_think tags are ignored. The model card also suggests different sampling settings for each mode: temperature 0.6 with top-p 0.95 when thinking, and 0.7 with top-p 0.8 when not, with top-k 20 in both. It warns against greedy decoding (always picking the single most likely word) in thinking mode, because it can cause endless repetition.

Which Qwen3 size should you run locally?

Qwen3 comes in six regular ("dense") sizes and two mixture-of-experts models. A mixture-of-experts model is split into many smaller specialist networks, and only a few of them work on each word, so it runs faster than its total size suggests. Its whole file still has to be loaded into memory.

The context window is how much text, counted in tokens (small chunks of words), the model can consider at once. For the 8B model, the card gives 32,768 tokens natively and 131,072 with an extension method called YaRN. The file sizes are estimates from simple arithmetic, using about 4.5 bits per parameter for a common 4-bit format; real files vary a little.

A practical guide: on a laptop with 8 GB of memory, start with the 4B model. With 16 GB, the 8B model is the natural choice and is a capable everyday assistant. With 32 GB or more, try the 14B, or the 30B-A3B if you want speed; the Qwen team says the 30B-A3B beats their earlier QwQ-32B while using a tenth of the active parameters. The 235B model is for servers, not personal computers. Whatever you choose, leave a few gigabytes free for the app and the conversation, and remember that thinking mode produces many more tokens, so it is slower on smaller machines.