Run Your First Model: Ollama Run Tutorial and Chat Tips

Start a conversation

ollama run qwen3:8b

The model downloads automatically the first time. Then you’ll see a prompt:

>>> Send a message (/? for help)

Type your question and press Enter. Press Ctrl + D or enter /bye to end the conversation.

Names and tags

The format of qwen3:8b is like a Docker image: the text before the colon is the model name, and the text after it is the tag (usually representing the parameter count or quantization version). If you omit the tag, Ollama uses latest, which usually points to the model’s recommended size.

ollama run gemma3        # gemma3:latest
ollama run gemma3:1b     # 1B small model, fast but less capable
ollama run gemma3:12b    # 12B, requires more memory

For help choosing, see Choosing a model and hardware requirements.

Ask a one-off question

Put your question directly after the command. The model exits after answering, which is useful in scripts:

ollama run qwen3:8b "What are five places in Taipei to visit on a rainy day?"

Pipe a file as input:

cat meeting-notes.txt | ollama run qwen3:8b "Summarize these meeting notes in five bullet points"

Enter multiline input

Wrap your input in three double quotes, """:

>>> """
... Please translate the following text into English:
... The weather is beautiful today,
... so let’s take a walk in the park.
... """

Interactive commands

At the >>> prompt, entries that start with / are Ollama commands and aren’t sent to the model:

Check generation speed

You can also add --verbose directly to the command. Here’s actual output from running qwen3:0.6b with Ollama 0.40 on a small server with no graphics card and only a single CPU core:

$ ollama run qwen3:0.6b --think=false --verbose "Introduce Taipei in one sentence"
Taipei is the capital of Taiwan, located in the Middle East Pacific region, in the northeast, and is an important economic and cultural center of Taiwan.

total duration:       46.080188811s
load duration:        33.014530786s
prompt eval count:    21 token(s)
prompt eval duration: 8.75411s
prompt eval rate:     2.40 tokens/s
eval count:           27 token(s)
eval duration:        4.179069999s
eval rate:            6.46 tokens/s

For comfortable reading, eval rate generally needs to be around 10 tokens per second or higher. This example gets only 6.46, and loading alone takes 33 seconds. That’s the cost of running without a GPU.

Also pay attention to the answer: “located in the Middle East Pacific region” is wrong. Smaller models are more likely to confidently make things up. A tiny model like 0.6B is suitable only for testing or very simple tasks. For real use, choose a larger model suited to your hardware.

Get answers in Traditional Chinese

Many models default to Simplified Chinese when you ask a question in Chinese. The most effective approach is to set a system prompt:

>>> /set system "You are a helpful assistant. Always respond in Traditional Chinese as commonly used in Taiwan."

To make this permanent, use a Modelfile to create a model that defaults to Traditional Chinese. Models with stronger Chinese-language ability, such as the Qwen family, also tend to perform better.

Thinking mode (reasoning models)

Models such as DeepSeek-R1 and Qwen3 think before answering. Their thinking appears before the answer. You can control this in the command:

ollama run qwen3:8b --think=false "What is 1+1?"   # Disable thinking for a faster answer
ollama run qwen3:8b --hidethinking "..."            # Think without showing the process

In interactive mode, use /set think and /set nothink to switch modes. (Available options may vary slightly across Ollama versions; enter /? to check.)

Read images (vision models)

Vision-capable models such as gemma3, qwen3.5, and llava can read images. Include the image path in your prompt:

ollama run gemma3 "What’s in this picture? ./photo.jpg"

On Windows, you can write the path as C:\Users\you\Pictures\photo.jpg.

Keep a model loaded in the background and stop it

By default, a model stays in memory for five minutes after its last use, so you don’t have to reload it for your next question.

ollama ps              # List loaded models
ollama stop qwen3:8b   # Unload immediately from memory

Tips

• The more specific your question, the better the answer. For example: “Explain it in three points, each no longer than 20 words.”
• Small models can confidently make things up. Verify important information yourself.
• When a conversation gets too long, the model may forget earlier content because of its context length. Use /clear to start over.

The next lesson shows you how to choose a model based on your computer’s specifications.

Updated October 12, 2026