aria-modelswitcher
Switch between Ollama models at runtime without restarting containers or reconfiguring LiteLLM/OpenClaw.
Why?
- •GLM-4.7-Flash-REAP - Smarter for complex text reasoning
- •Qwen3-VL - Has vision capabilities for image tasks
- •Both run locally on Mac Metal GPU via Ollama
Model Aliases
| Alias | Full Model Name | Use Case |
|---|---|---|
glm | hf.co/unsloth/GLM-4.7-Flash-REAP-23B-A3B-GGUF:Q3_K_S | Default. Smart text reasoning |
qwen3-vl | qwen3-vl:8b | Vision/image tasks |
qwen2.5 | qwen2.5:7b | Backup text model |
Usage
bash
# List available models
exec python3 /root/.openclaw/workspace/skills/run_skill.py model_switcher list_models '{}'
# Switch to GLM for text tasks (default)
exec python3 /root/.openclaw/workspace/skills/run_skill.py model_switcher switch_model '{"model": "glm"}'
# Switch to Qwen3-VL for vision/image analysis
exec python3 /root/.openclaw/workspace/skills/run_skill.py model_switcher switch_model '{"model": "qwen3-vl"}'
# Check current model
exec python3 /root/.openclaw/workspace/skills/run_skill.py model_switcher get_current_model '{}'
# Pull a model if not available
exec python3 /root/.openclaw/workspace/skills/run_skill.py model_switcher pull_model '{"model": "qwen3-vl"}'
How It Works
- •Model preference is stored in
/root/.openclaw/workspace/memory/model_preference.json - •
OllamaSkillreads this file on each request to determine which model to use - •No container restart needed - changes take effect immediately
- •LiteLLM/OpenClaw are unaware - they just see "Ollama" as the provider
Recommended Workflow
- •Use GLM as default for most tasks (smarter, better reasoning)
- •Switch to Qwen3-VL when you need to analyze images
- •Switch back to GLM after vision tasks complete
Troubleshooting
Model not found
bash
# Pull the model first
exec python3 /root/.openclaw/workspace/skills/run_skill.py model_switcher pull_model '{"model": "glm"}'
Ollama not reachable
Check that Ollama is running natively on Mac (not in Docker) for Metal GPU acceleration:
bash
OLLAMA_HOST=0.0.0.0:11434 ollama serve