DeepSeek is a Chinese lab whose open-weight releases attracted unusual attention for matching far more expensive models on reasoning tasks. The locally runnable versions are distilled — smaller models trained on the larger one's behaviour — which is how a reasoning model fits on a laptop at all.
Install and run
# check the library for current tags first
ollama run deepseek-r1:7b
Smaller distilled variants exist, including a 1.5B, which is worth trying on a constrained machine.
What you will notice immediately
It thinks before answering, visibly. Ask it something non-trivial and you will see it working through the problem — considering an approach, rejecting it, trying another — before producing a final answer.
That reasoning text is the point, not a bug. The model is spending tokens to improve its answer, which is the same mechanism behind reasoning modes in hosted models.
Two practical consequences:
- It is slower, sometimes much slower, because it generates far more tokens per answer.
- You can read its working, which is genuinely useful — when the answer is wrong you can usually see exactly where it went wrong.
When to use it, and when not to
Worth the wait for: multi-step problems, logic puzzles, debugging, mathematics, anything where you want to check the reasoning rather than trust a conclusion.
Not worth it for: summarising, rewriting, extraction, classification, translation. For those you are paying a large time cost for thinking the task does not need. Use Mistral or a small Qwen instead.
This is the clearest case in the tutorial for having several models installed rather than looking for one winner. They are not competing for the same job.
A note on model origin
You are running the weights locally, so nothing you type is sent to the publisher, in China or anywhere else. That is worth stating plainly because it is a common and reasonable worry, and the answer is genuinely reassuring: a local model has no network connection to its creator.
What is worth knowing is that any model reflects its training data and the choices of whoever made it, which shows up in how it handles politically sensitive subjects. That is true of every model from every country. If you are building something where that matters, test it on the specific topics you care about rather than assuming.