Ollama Hub

Ollama Hub

Manage and benchmark local Ollama models.

Use when listing models, pulling new ones, benchmarking performance, or comparing models.

Requirements

• Ollama installed and running (ollama serve or systemd service)
• No API keys needed

Instructions

1. List installed models:

   ollama list                    # name, size, modified date
   ollama show <model>            # detailed info (parameters, template, license)
   

2. Pull / remove models:

   ollama pull llama3.3:70b       # download a model
   ollama pull mistral:latest     # latest version
   ollama rm <model>              # remove (confirm with user first!)
   

3. Benchmark a model:

   # Time a response
   time ollama run <model> "Explain quantum computing in 3 sentences" --verbose 2>&1
   
   # Extract tokens/sec from verbose output
   ollama run <model> "Hello" --verbose 2>&1 | grep "eval rate"
   

4. Compare models — run same prompt across multiple models:

   ## 📊 Ollama Model Benchmark
   **Prompt:** "Explain quantum computing in 3 sentences"
   **Hardware:** [CPU/GPU specs]
   
   | Model | Size | Tokens/sec | Response Time | Quality |
   |-------|------|-----------|--------------|---------|
   | llama3.3:8b | 4.7GB | 42 t/s | 2.1s | ⭐⭐⭐⭐ |
   | mistral:7b | 4.1GB | 48 t/s | 1.8s | ⭐⭐⭐ |
   | phi3:mini | 2.3GB | 65 t/s | 1.2s | ⭐⭐⭐ |
   

5. Check Ollama status:

   curl -s http://localhost:11434/api/tags | jq .    # API check
   systemctl status ollama                            # service status
   ollama ps                                          # running models
   

Model Naming

Format: name:tag — e.g., llama3.3:8b, mistral:latest, codellama:13b-instruct

Common tags: latest, 7b, 13b, 70b, instruct, code

Edge Cases

• Ollama not running: Start with ollama serve or systemctl start ollama.
• Insufficient disk space: Check df -h before pulling large models. 70B models need ~40GB.
• Insufficient RAM: Models need RAM ≈ model size. 7B ≈ 8GB RAM, 70B ≈ 48GB RAM.
• GPU vs CPU: Performance varies dramatically. Note hardware in benchmarks.
• Model not found: Check spelling. Use ollama list to see available names. Search at ollama.com/library.
• Slow downloads: Large models take time. Use ollama pull with patience; it supports resume.

Troubleshooting

• Port 11434 in use: Another Ollama instance may be running. lsof -i :11434.
• CUDA errors: Check GPU drivers with nvidia-smi. Reinstall Ollama if needed.
• Model corrupted: Remove and re-pull: ollama rm <model> && ollama pull <model>.