Picking and Pulling Models
Reading Ollama's Model Library
The model library at https://ollama.com/library (or the search page at https://ollama.com/search) is just a browsable index. Every model you see there can be pulled with ollama pull $MODEL and run with ollama run $MODEL. What matters for picking one is reading the metadata correctly: the capability tags, the size variants, and the specific tag you append after the colon.
Anatomy of a Model Entry
A typical card on the library page looks like this:
Model card for qwen3.6
Three things to read:
Capability tags (
vision,tools,thinking,audio,cloud,embedding) tell you what the model supports.Size variants (
27b,35b,e2b,e4b) are the parameter counts you can pull.Tags count (
22 Tags) is the number of specific variants available, including quantizations and fine-tunes. For example, for qwen3.6, click through to https://ollama.com/library/qwen3.6/tags to see all of them with their disk size and context window.
Capability Tags
These are the labels that determine what the model can actually do. Filter the library by them:
# For vision and tools:
https://ollama.com/search?c=vision&c=tools
# For embeddings
https://ollama.com/search?c=embedding
# For cloud models
https://ollama.com/search?c=cloud
# etc
Follow the table below to understand the meaning of each tag.
| Tag | What it means | What you do with it |
|---|---|---|
vision | Accepts image input alongside text | Pass images via the API or paste them into a compatible UI |
tools | Supports function/tool calling with structured arguments | Build agents, use with frameworks like LangChain, integrate with MCP |
thinking | Emits a reasoning trace before the answer | Toggle with the think parameter; better on math, code, multi-step tasks |
audio | Accepts audio input | Transcription, speech understanding |
embedding | Produces vectors, not chat output | Use the /api/embed endpoint for RAG, semantic search |
cloud | Runs on Ollama's hosted infrastructure, not your machine | For models too big for local hardware; requires ollama signin |
A model can carry several at once. qwen3.5 is tagged vision, tools, thinking, cloud, meaning the same model family supports images, tool calls, reasoning, and has a cloud variant for the largest sizes.
The cloud tag can be a bit confusing. A model like kimi-k2.6 listed only as cloud is not
Local AI Engineering with Ollama
Run, understand, customize, fine-tune, and build agentic apps on your own hardwareEnroll now to unlock all content and receive all future updates for free.