Mistral is a French lab whose open-weight models built a reputation for doing more with fewer parameters. If an 8B Llama is marginal on your hardware, Mistral is the obvious next thing to try.
Install and run
Check the tag on ollama.com/library as before, then:
ollama run mistral
The bare mistral tag resolves to the default build, which is a
7B class model — roughly 4 GB on disk and comfortable on a 16 GB machine.
Where it tends to do well
Treat the following as tendencies reported by users rather than benchmark claims, and test on your own work:
- Instruction following. It is often more literal about doing exactly what you asked, which matters when you want output in a specific format.
- European languages. Unsurprisingly strong on French, and generally solid across European languages.
- Terseness. It tends to pad less than some alternatives, which is a real advantage when you are extracting data rather than conversing.
Comparing two models properly
Now that you have two installed, the useful exercise is comparing them on your work rather than on someone's leaderboard. Run both against the same real task:
cat sample.txt | ollama run llama3.1:8b "Extract every date and what happened on it, as a list"
cat sample.txt | ollama run mistral "Extract every date and what happened on it, as a list"
Public benchmarks measure average performance across generic tasks. You care about one specific task with your own data, and the two frequently disagree. Ten minutes of this is worth more than any comparison article.
Managing what you have downloaded
Two models in, disk use becomes real:
# what you have and how large it is
ollama list
# remove one you have finished with
ollama rm mistral
# details of a specific model
ollama show mistral
ollama show is worth running once. It prints the parameter
count, quantisation and context length of what you actually downloaded, which
is the ground truth when a tag is ambiguous.