Llama is Meta's open-weight family and the most widely used starting point, largely because almost every tool supports it.
Check the current tag first
This is a habit worth forming now, because it is what stops a tutorial going
stale on you. Open
ollama.com/library
and search for llama. The library lists each available version and
size with its exact tag.
Tags look like llama3.1:8b — family, version, parameter count.
Use what the library shows today rather than what any article says, including
this one.
Pull and run
Pick the size from chapter two's table. For a 16 GB machine, an 8B model is the right first choice:
ollama run llama3.1:8b
The first run downloads several gigabytes, then drops you into a prompt. Subsequent runs start in seconds.
If you only want the download without a chat session:
ollama pull llama3.1:8b
Using the prompt
Type a question and press Enter. A few things worth knowing:
/bye— exit the session./clear— wipe the conversation context and start fresh./?— list the available commands.- Triple quotes
"""open and close a multi-line input, which you need for pasting documents or code.
Try something that shows what it is good at rather than what it is not:
Summarise the following in three bullet points: """
...paste a few paragraphs here...
"""
One-shot use from the command line
You do not have to enter the chat at all:
ollama run llama3.1:8b "Rewrite this in plain English: the aforementioned party shall be liable"
This is the form you will use in scripts. It reads from standard input too, which makes it pipeable:
cat notes.txt | ollama run llama3.1:8b "Summarise these meeting notes as action items"
If it is too slow
Three things to try, in order:
- Drop a size. An 8B struggling on your machine becomes a 3B that works. Capability loss is smaller than you expect for summarising and extraction.
- Close other applications. Browsers are the usual culprit; memory pressure forces swapping and generation collapses.
- Check you are not swapping. If the model nearly fits, the operating system pages to disk and speed falls off a cliff rather than degrading gently.
If an 8B model will not run comfortably, that is information rather than failure — your machine is a 3B machine for now, and chapter eight explains how quantisation buys back some headroom.