# Model Storage Directory

This directory stores GGUF quantized model weights and vision encoders (`mmproj`).
Model binary weights (`*.gguf`) are excluded from Git version control via `.gitignore`.

## Expected Directory Layout

```
models/
├── Qwen3.5/
│   ├── Qwen3.5-4B-Q5_K_M.gguf                     # Instruct / Worker Agent
│   ├── mmproj-BF16.gguf                          # Multimodal Vision Projector
│   └── chat_template.jinja -> ../../templates/Qwen3.5/chat_template.jinja
├── Qwen3.8/
│   ├── Qwen3.8-27B-UD-Q4_K_XL.gguf               # Reasoning / Planning Agent
│   ├── SuperQwen3.8-27b-abliterated-Q4_K_M.gguf  # Alternative 4-bit quant
│   ├── mmproj-SuperQwen3.8-27b-abliterated-Q8_0.gguf
│   └── chat_template.jinja -> ../../templates/Qwen3.8/chat_template.jinja
└── muse-glimmer/
    ├── Muse-Glimmer-30B-UD-Q4_K_XL.gguf          # Deep Reasoning Solo Model
    ├── mmproj-Muse-Glimmer-30B-Q8_0.gguf         # Vision Projector
    └── chat_template.jinja -> ../../templates/muse-glimmer/chat_template.jinja
```

## Recommended Quantizations & Sourcing

| Model Name | Quantization | Size | Purpose | Recommended Port |
| :--- | :--- | :--- | :--- | :--- |
| **SuperQwen 3.8 27B** | `UD-Q4_K_XL` or `Q4_K_M` | ~17-18 GB | Planner / Architect agent | `11434` |
| **Qwen 3.5 4B** | `Q5_K_M` | ~3.9 GB | Instruct / Implementation agent | `11435`, `11436` |
| **Muse-Glimmer 30B** | `UD-Q4_K_XL` | ~17 GB | Deep reasoning solo agent | `11434` |

All chat templates are tracked in `templates/` and referenced automatically by `bin/llm-runner`.