/suggest-ollama-models

You are helping the user review their Ollama models and suggest new ones based on their hardware.

Process

1. Check currently installed models

   • Run: ollama list
   • Show model sizes and last modified dates
   • Calculate total disk usage

2. Assess hardware capabilities

   • Check GPU VRAM: rocm-smi (for AMD) or nvidia-smi (for NVIDIA)
   • Check system RAM: free -h
   • Determine recommended model sizes:
     • < 8GB VRAM: 7B models and smaller
     • 8-16GB VRAM: up to 13B models
     • 16-24GB VRAM: up to 34B models
     • 24GB+ VRAM: 70B+ models possible

3. Identify user's needs

   • Ask about use cases:
     • General chat
     • Code generation
     • Data analysis
     • Creative writing
     • Vision/multimodal
     • Specialized domains

4. Suggest models by category

   General Purpose:

   • llama3.2 (3B, 8B)
   • qwen2.5 (7B, 14B, 32B)
   • mistral (7B)
   • gemma2 (9B, 27B)

   Code:

   • codellama (7B, 13B, 34B)
   • deepseek-coder (6.7B, 33B)
   • starcoder2 (7B, 15B)

   Fast/Small:

   • tinyllama (1.1B)
   • phi3 (3.8B)

   Multimodal:

   • llava (7B, 13B, 34B)
   • bakllava (7B)

   Specialized:

   • meditron (medical)
   • sqlcoder (SQL generation)
   • wizardmath (mathematics)

5. Consider quantization levels

   • Explain different quants (Q4, Q5, Q8, etc.)
   • Suggest appropriate quant for their VRAM

6. Cleanup suggestions

   • Identify duplicate models
   • Suggest removing unused models: ollama rm <model>
   • Free up space for new models

Output

Provide a report showing:

• Currently installed models and total size
• Hardware capacity summary
• Recommended models based on:
  • Available VRAM
  • User's use cases
  • Current gaps in model coverage
• Commands to install suggested models
• Models that could be removed to save space