返回 Skill 列表
extension
分类: 开发与工程无需 API Key

minion-models

管理Ollama模型以供下属使用。拉取新模型,列出可用模型,切换预设,检查磁盘使用情况。在设置或更改模型时使用。

person作者: jakexiaohubgithub

Minion Models

Manage Ollama models for your minion squad.

Quick commands

List installed models

ollama list

Pull a model

ollama pull qwen2.5-coder:1.5b

Check model info

ollama show qwen2.5-coder:1.5b

Remove a model

ollama rm qwen2.5-coder:0.5b

Presets

| Preset | Models | Download | RAM | |--------|--------|----------|-----| | nano | qwen2.5-coder:0.5b | ~350MB | ~1GB | | small | qwen2.5-coder:1.5b | ~1GB | ~2GB | | medium | qwen2.5-coder:7b | ~4.5GB | ~8GB | | large | qwen2.5-coder:14b | ~9GB | ~16GB |

Pull preset models

nano:

ollama pull qwen2.5-coder:0.5b

small (recommended):

ollama pull qwen2.5-coder:1.5b

medium:

ollama pull qwen2.5-coder:7b

large:

ollama pull qwen2.5-coder:14b

Switch preset

Edit llm_gc/config/models.yaml and change the preset line:

preset: small  # Change to: nano, small, medium, or large

Or use sed:

sed -i.bak 's/^preset:.*/preset: medium/' llm_gc/config/models.yaml

Check disk usage

# Total Ollama storage
du -sh ~/.ollama/models

# Per-model breakdown
ls -lh ~/.ollama/models/blobs/ | head -20

Recommended models

| Task | Model | Why | |------|-------|-----| | Quick patches | qwen2.5-coder:1.5b | Fast, good enough | | Quality patches | qwen2.5-coder:7b | Better reasoning | | Code review | qwen2.5-coder:7b+ | Needs context | | Simple questions | qwen2.5-coder:0.5b | Speed matters |

Alternative models

# DeepSeek (alternative to Qwen)
ollama pull deepseek-coder:1.3b
ollama pull deepseek-coder:6.7b

# CodeLlama (Meta)
ollama pull codellama:7b

# StarCoder
ollama pull starcoder2:3b

Troubleshooting

Model not found:

ollama pull <model-name>

Slow responses:

  • Try smaller model
  • Check htop for RAM pressure
  • Reduce --workers in swarm

Out of disk space:

# Remove unused models
ollama rm <model-name>

# Check what's installed
ollama list

Model quality issues:

  • Upgrade preset: nano → small → medium
  • Add more context with --read
  • Simplify the task