Ollama Deployment & Multi-Model Management

Use this skill when the user asks to install, manage, or test multiple local LLMs via Ollama on a server or workstation (CPU or GPU).

When to use

• User wants to “install several models” or “set up a local LLM system”
• User asks about running models on CPU without GPU
• You need to manage an Ollama inventory (list, pull, remove, switch models)
• User wants to identify which models support agentic/tool-calling workflows
• Resource planning: checking RAM/disk before downloading models

Core workflow

1. Check resources first

free -h                     # RAM + swap
nproc                       # CPU cores
df -h /                     # Disk space
ollama list                 # Already installed models

2. Batch pull strategy

Ollama downloads sequentially per process. Launch background pulls in parallel with ollama pull <model> in separate terminal sessions. Monitor with ollama list or check process output.

Key insight: Launch ALL pulls at once in background — Ollama handles its own queue. This is faster than waiting for each.

3. Model naming pitfalls

Always verify model tag on ollama.com/library before pulling if unsure.

4. Agentic model identification

Models with native function calling / tool use support (best for agent workflows):

• gemma4:e2b — Google DeepMind, native function calling, vision, reasoning, 128K ctx. Best agentic edge model.
• mistral:7b — Native function calling, fastest CPU agent (25+ tok/s).
• phi4-mini — Microsoft, 128K ctx, reasoning-dense, excellent lightweight agent.
• qwen2.5:7b-instruct — Good tool calling, multilingual, code.
• llama3.2 / llama3.1:8b — Largest ecosystem, solid instruction following.
• deepseek-r1:8b — Excellent reasoning but limited tool calling.

5. CPU-only model tier guide

Based on benchmarks (Q4_K_M quantization, ~6-core CPU):

First launch is slow (5-15s load time into RAM). Subsequent calls within same session are instant if RAM not reclaimed.

6. Switching models on a server

For production use where you switch models without reloading:

• Run ollama serve in the background
• Use the OpenAI-compatible API at http://localhost:11434/v1
• Just change the model field in requests — Ollama hot-swaps

7. Disk space estimation

Each model at Q4 quantization:

• 0.6B-1B: ~500MB-1GB
• 2-3B: ~1.5-2.5GB
• 7-8B: ~4-5GB
• Gemma 4 E2B: ~7.2GB
• Gemma 4 E4B: ~9.6GB

Add ~10% overhead for blobs/manifest metadata.

Verification

After bulk install:

# List all models with sizes
ollama list

# Quick smoke test (smallest model first — loads fastest)
ollama run qwen3:0.6b "Hello in 3 words"

# Check API is responding
curl -s http://localhost:11434/api/tags | head -5

References

• model-library.md — Full Ollama model catalog with agentic annotations, CPU benchmarks, and deployment notes
• troubleshooting.md — Common Ollama issues: disk space, pull failures, version compatibility, model not found

Resources

• Ollama library: https://ollama.com/library
• Ollama GitHub: https://github.com/ollama/ollama
• Install: curl -fsSL https://ollama.com/install.sh | sh
• API docs: https://github.com/ollama/ollama/blob/main/docs/api.md