Models, tags and sizes

How model names and tags work, what the size in a tag means, and which kind of model fits which job.

Also called model:tag, latest

What: A model reference is name:tag, such as qwen3:8b or gemma3:12b-it-q8_0. The tag picks the size and quantisation; leaving it out means latest, a sensible default variant.
Why: Pinning an explicit tag makes results reproducible and stops latest changing under you.

Watch out

Assuming latest means the biggest or best version; it is usually a mid-sized 4-bit build.

Also called 1b, 8b, 70b

What: The number in a tag, such as 8b, is billions of weights. More parameters usually mean better knowledge and reasoning, and proportionally more memory and slower generation.
Why: It is the first filter when choosing: find the largest model your memory fits at a decent quantisation.
Example: Rough guide at 4-bit: 1 to 4B runs on almost anything, 7 to 14B suits a 16 GB machine, 27 to 32B wants 24 GB or more, 70B needs around 48 GB.

Also called chat model, it

What: Instruct (chat) models are fine-tuned to follow instructions in a conversation. Base models only continue text. Library defaults are almost always instruct variants.
Why: A base model given a question may simply write more questions; use instruct models for apps.

What: Popular open-weight families include Llama, Qwen, Gemma, Mistral, Phi, DeepSeek and gpt-oss, each in several sizes, with specialised coder, vision and embedding variants.
Why: Families differ in languages, coding skill, tool-calling reliability and licence, so test two or three on your own prompts.

Watch out

Choosing by leaderboard rank alone; your task and your hardware matter more.

Also called reasoning models, think

What: Models such as Qwen3, DeepSeek-R1 and gpt-oss can reason step by step before answering. Ollama returns the reasoning separately, and the think option turns it on or off.
Why: Thinking helps maths, code and planning, but costs many extra tokens and seconds; turn it off for simple chat.

Also called multimodal

What: Models such as Gemma 3 and Qwen-VL variants accept images alongside text, through the images field in the API or a file path in the CLI prompt.
Why: Screenshots, receipts and diagrams can be described or extracted locally without uploading them anywhere.

Shell

ollama run gemma3 "What does this chart show? ./sales.png"

Also called nomic-embed-text, mxbai-embed-large

What: Small models such as nomic-embed-text or mxbai-embed-large that turn text into vectors instead of generating text. They are used through the embed API, not ollama run.
Why: They power local semantic search and RAG with no external API.

Also called hf.co

What: Any GGUF repository on Hugging Face can be run directly with hf.co/user/repo, optionally with a quantisation tag.
Why: You get fine-tunes and new releases that are not in the Ollama library yet.

Shell

ollama run hf.co/bartowski/Llama-3.2-3B-Instruct-GGUF:Q4_K_M
ollama pull hf.co/user/repo:Q8_0

Go deeper