docs/HARDWARE_PROFILE_JSON.md

# Hardware Profile JSON

JackAILocal SaaS Builder accepts an optional hardware profile JSON object.

If the field is empty, the selected sample profile is used.

## Minimum fields

```json
{
  "os": "windows",
  "cpu_threads": 8,
  "ram_gb": 16,
  "vram_gb": 0,
  "gpu_name": "CPU-only office laptop",
  "disk": "USB 3.2 SSD",
  "usb_storage_gb": 128
}
```

## RTX 3090 example

```json
{
  "os": "windows",
  "cpu_threads": 24,
  "ram_gb": 64,
  "vram_gb": 24,
  "gpu_name": "NVIDIA RTX 3090",
  "disk": "USB-C NVMe",
  "usb_storage_gb": 512
}
```

## Optional whichllm block

```json
{
  "os": "windows",
  "cpu_threads": 24,
  "ram_gb": 64,
  "vram_gb": 24,
  "gpu_name": "NVIDIA RTX 3090",
  "disk": "USB-C NVMe",
  "usb_storage_gb": 512,
  "whichllm": {
    "source": "optional",
    "top_model": "Qwen/Qwen3.6-27B"
  }
}
```

## Selection behavior

The current hackathon Gradio app is rule-based. It does not perform a live hardware scan inside the browser.

The selected sample profile or pasted JSON is normalized, then compared against the signed model catalog.

The standard pack defaults to a practical laptop baseline up to 9B. Power profiles can select compatible models up to the hackathon limit of 32B.