Qwen is Alibaba's open-weight family. Its distinguishing feature is range: it is published in many more sizes than most families, from fractions of a billion parameters up to very large models.
That matters more than it sounds. It means there is almost certainly a Qwen that fits your exact machine, rather than forcing the usual choice between "too small to be useful" and "too big to run".
Install and run
Check the library for current tags, then pick by your memory budget:
# small, for 8 GB machines or fast batch work
ollama run qwen2.5:1.5b
# the usual starting point on 16 GB
ollama run qwen2.5:7b
Larger variants exist for machines with more headroom. Chapter two's table tells you what you can hold.
Where it is strongest
- Multilingual work, and Chinese in particular, where it is generally considered the strongest of the widely available open families.
- Coding, especially in the variants published specifically for it — look for a coder tag in the library.
- Small-size performance. The 1.5B models are genuinely usable for classification and extraction, which is not true of every family at that size.
Why small models deserve more attention
There is a reflex to always run the largest model that fits. It is often wrong.
For a task like "does this support ticket mention a refund request, yes or no", a 1.5B model answers correctly and returns in a fraction of the time. Run that over fifty thousand tickets and the difference between a 1.5B and a 7B is hours against days.
The right question is not "what is the best model I can run" but "what is the smallest model that does this task reliably". Test the small one first and move up only when it fails.
Measuring the difference
Ollama can show you the numbers:
ollama run qwen2.5:1.5b --verbose "Classify the sentiment of: the delivery was late but the staff were kind"
The --verbose flag prints timings and tokens per second after
the response. Run it against both the 1.5B and the 7B on the same prompt and
you will have a real number for the trade-off on your hardware, rather than an
impression.