ollama

Files (1)

• SKILL.md

SKILL.md (first lines)

# ollama

Run and manage LLM models using `ollama`, either locally or on a remote machine
accessed directly over the network.

## Orientation

```bash
which ollama
ollama --version
ollama --help
```

## Workflow

### Local

1. Confirm the tool is present: `ollama --version`
2. Start the server if not running: `ollama serve`
3. Pull a model: `ollama pull <model>`
4. Run a model interactively: `ollama run <model>`
5. List downloaded models: `ollama list`
6. Remove a model: `ollama rm <model>`

### Remote

Ollama has no built-in remote relay (unlike LM Studio's LM Link). Access a
remote Ollama instance by pointing directly at its host and port over a secure
network (e.g. Tailscale):

```bash
# List available models on a remote instance
curl http://<remote-host>:11434/api/tags

# Run inference (non-streaming)
curl http://<remote-host>:11434/api/chat \
  -d '{"model":"<model>","messages":[{"role":"user","content":"Hello"}],"stream":false}'

# OpenAI-compatible endpoint — works with most tools
# base_url: http://<remote-host>:11434/v1
# model: <model-name>
```

Replace `<remote-host>` with the remote machine's Tailscale IP or hostname.
Port 11434 is the Ollama default.

## Discovering models on ollama.com

Browse and search at https://ollama.com/search — models are filterable by
category (e.g. vision, tools, embedding, code).

Each model has a library page at `https://ollama.com/library/<model>` listing
all available tags (sizes and quantizations).

Common tag suffixes:

| Tag | Meaning |
|---|---|
| `:latest` | Default tag — usually the recommended size |
| `:1b`, `:3b`, `:7b` | Parameter count (smaller = faster, less capable) |
| `q4_0`, `q4_k_m`, `q8_0` | Quantization level (higher = more accurate, larger) |
| `fp16` | Full precision — largest and most accurate |

If no tag is specified, `ollama pull <model>` fetches `:latest`.

### Inspect a model on ollama.com without pulling it

Use `https://ollama.com/api/show` to fetch model metadata directly from the
ollama.com library — no local Ollama install required:

```bash
# Show metadata for a model (architecture, context length, capabilities, etc.)
curl -s -X POST https://ollama.com/api/show \
  -d '{"name":"llama3.2:latest"}' | jq .

# Example output fields: details.family, details.parameter_size,
# model_info.*.context_length, capabilities (completion, vision, tools, etc.)
```

The `name` field must match a tag listed on the model's library page
(`https://ollama.com/library/<model>`). Use `:latest` if unsure of the tag.

### Query ollama.com featured models with jq

```bash
# List all featured model names
curl -s https://ollama.com/api/tags | jq '[.models[].name]'

# Show name and size (in GB), sorted smallest first
curl -s https://ollama.com/api/tags | jq '[.models[] | {name, size_gb: (.size/1e9 | . * 10 | round / 10)}] | sort_by(.size_gb)'

# Filter to models under a size threshold (e.g. under 10 GB)
curl -s https://ollama.com/api/tags | jq '[.models[] | select(.size < 10000000000) | {name, size_gb: (.size/1e9 | . * 10 | round / 10)}]'
```

Note: `https://ollama.com/api/tags` returns a curated/featured subset (~40
models), not the full library. Browse the full catalog at
`https://ollama.com/search`.

### Pull and inspect a model

```bash
# Pull a specific model and tag
ollama pull <model>:<tag>

# Show model info (architecture, context length, quantization, capabilities)
ollama show <model>:<tag>

# Show individual details
ollama show <model>:<tag> --modelfile     # Modelfile used to create it
ollama show <model>:<tag> --parameters   # Runtime parameters (temperature, stop tokens, etc.)
ollama show <model>:<tag> --template     # Prompt template
ollama show <model>:<tag> --system       # System prompt (if any)
ollama show <model>:<tag> --license      # License
ollama show <model>:<tag> --verbose      # Full detail dump

# List all locally pulled models
ollama list

From jason-riddle/agent-skills, MIT.

More from jason-riddle/agent-skills