Choosing Models from the Library

Tags encode size and quantization

The Ollama library lists models by family, and each model has many tags. A tag usually combines the parameter count and sometimes the variant and quantization, for example qwen2.5:7b, qwen2.5:7b-instruct-q8_0 or gemma3:4b. The default tag for most models is an instruct variant at a 4-bit quantization, which is a sensible starting point. Choose by task first: general chat, coding (models with coder in the name), embeddings (models such as nomic-embed-text that return vectors, not text), or vision (models that accept images). Then choose by hardware: the size shown on the tag page must fit in your VRAM or RAM with headroom for context. Ollama can also run GGUF models straight from Hugging Face with ollama run hf.co/<user>/<repo>, optionally followed by a quantization tag, which opens up thousands of community quantizations.

Picking a phone variant

A phone model comes as 6/128 or 8/256. Same phone, different memory configurations and prices. A model tag is the same: the family is the phone, the tag tells you the size and how tightly it is packed, and you buy the one your pocket (your RAM) can afford.

Pulling specific tags

Be explicit about tags in scripts so everyone on the team runs the same thing.

ollama pull qwen2.5:7b                 # default quant for the 7B size
ollama pull nomic-embed-text           # an embedding model
# Run a GGUF straight from Hugging Face, choosing a quant
ollama run hf.co/bartowski/Llama-3.2-3B-Instruct-GGUF:Q4_K_M

त्वरित जाँच: You need vectors for semantic search, not chat replies. What kind of model should you pull?

• A coder model
• An embedding model such as nomic-embed-text
• The largest instruct model available
• A base model