[Bug] Foundry does not show any NPU models, only shows GPU and CPU models

@bmehta001 is already working on this.

Since Aug 12, 2026.

### Describe the issue

I am using an HP EliteBook 8 G1a - Ryzen 5 230 / 760m graphics / NPU

When attempting to find models to load with `foundry model list` I can see only CPU and GPU models. No NPU models are available.

### To reproduce

Downloaded and install foundry-0.10.3-win-x64-winml.msix

run `foundry model list`

Output:
```
foundry model list
╭───────────────────────────────────┬────────────┬─────────┬────────┬───────┬────────╮
│ Model Name │ Type │ Size │ Device │ Tools │ Cached │
├───────────────────────────────────┼────────────┼─────────┼────────┼───────┼────────┤
│ deepseek-r1-14b │ Chat │ 10.3 GB │ GPU │ ○ │ ○ │
│ deepseek-r1-7b │ Chat │ 5.6 GB │ GPU │ ○ │ ○ │
│ gpt-oss-20b │ Chat │ 11.8 GB │ GPU │ ○ │ ○ │
│ ministral-3-3b-instruct-2512 │ Multimodal │ 3.6 GB │ GPU │ ● │ ○ │
│ mistral-7b-v0.2 │ Chat │ 4.1 GB │ GPU │ ○ │ ○ │
│ mistral-nemo-12b-instruct │ Chat │ 7.3 GB │ GPU │ ● │ ○ │
│ nemotron-3.5-asr-streaming-0.6b │ Speech │ 756 MB │ CPU │ ○ │ ○ │
│ nemotron-speech-streaming-en-0.6b │ Speech │ 696 MB │ CPU │ ○ │ ○ │
│ nemotron-speech-streaming-es-0.6b │ Speech │ 729 MB │ CPU │ ○ │ ○ │
│ olmo-3-7b-instruct │ Chat │ 5.5 GB │ GPU │ ● │ ○ │
│ parakeet-tdt-0.6b-v2 │ Speech │ 692 MB │ CPU │ ○ │ ○ │
│ phi-3-mini-128k │ Chat │ 2.1 GB │ GPU │ ○ │ ○ │
│ phi-3-mini-4k │ Chat │ 2.1 GB │ GPU │ ○ │ ○ │
│ phi-3.5-mini │ Chat │ 2.2 GB │ GPU │ ○ │ ○ │
│ phi-4 │ Chat │ 8.4 GB │ GPU │ ○ │ ○ │
│ phi-4-mini │ Chat │ 3.7 GB │ GPU │ ● │ ○ │
│ phi-4-mini-reasoning │ Chat │ 3.1 GB │ GPU │ ○ │ ○ │
│ phi-4-reasoning │ Chat │ 8.4 GB │ GPU │ ○ │ ○ │
│ qwen2.5-0.5b │ Chat │ 700 MB │ GPU │ ● │ ○ │
│ qwen2.5-1.5b │ Chat │ 1.5 GB │ GPU │ ● │ ○ │
│ qwen2.5-14b │ Chat │ 9.3 GB │ GPU │ ● │ ○ │
│ qwen2.5-7b │ Chat │ 5.2 GB │ GPU │ ● │ ○ │
│ qwen2.5-coder-0.5b │ Chat │ 528 MB │ GPU │ ● │ ○ │
│ qwen2.5-coder-1.5b │ Chat │ 1.3 GB │ GPU │ ● │ ○ │
│ qwen2.5-coder-14b │ Chat │ 8.8 GB │ GPU │ ● │ ○ │
│ qwen2.5-coder-7b │ Chat │ 4.7 GB │ GPU │ ● │ ○ │
│ qwen3-0.6b │ Chat │ 529 MB │ GPU │ ● │ ○ │
│ qwen3-1.7b │ Chat │ 1.4 GB │ GPU │ ● │ ○ │
│ qwen3-14b │ Chat │ 9.9 GB │ GPU │ ● │ ○ │
│ qwen3-4b │ Chat │ 2.9 GB │ GPU │ ● │ ○ │
│ qwen3-8b │ Chat │ 6.0 GB │ GPU │ ● │ ○ │
│ qwen3-embedding-0.6b │ Embedding │ 515 MB │ GPU │ ○ │ ○ │
│ qwen3-embedding-8b │ Embedding │ 6.0 GB │ GPU │ ○ │ ○ │
│ qwen3-vl-2b-instruct │ Multimodal │ 1.3 GB │ CPU │ ● │ ○ │
│ qwen3-vl-4b-instruct │ Multimodal │ 2.7 GB │ CPU │ ● │ ○ │
│ qwen3-vl-8b-instruct │ Multimodal │ 5.0 GB │ CPU │ ● │ ○ │
│ qwen3.5-0.8b │ Multimodal │ 1.3 GB │ GPU │ ● │ ○ │
│ qwen3.5-2b │ Multimodal │ 3.1 GB │ GPU │ ● │ ○ │
│ qwen3.5-2b-text │ Chat │ 1.3 GB │ GPU │ ● │ ○ │
│ qwen3.5-4b │ Multimodal │ 5.4 GB │ GPU │ ● │ ○ │
│ qwen3.5-9b │ Multimodal │ 7.0 GB │ GPU │ ● │ ○ │
│ smollm3-3b │ Chat │ 2.2 GB │ GPU │ ● │ ○ │
│ whisper-base │ Speech │ 205 MB │ CPU │ ○ │ ○ │
│ whisper-large-v3-turbo │ Speech │ 1.3 GB │ CPU │ ○ │ ○ │
│ whisper-medium │ Speech │ 937 MB │ CPU │ ○ │ ○ │
│ whisper-small │ Speech │ 438 MB │ CPU │ ○ │ ○ │
│ whisper-tiny │ Speech │ 131 MB │ CPU │ ○ │ ○ │
╰───────────────────────────────────┴────────────┴─────────┴────────┴───────┴────────╯
Type: Chat / Multimodal / Speech / Embedding = model task.
Device: NPU / GPU / CPU = default hardware target selected for this machine.
Tools: supports tool calling; the model can request API calls or app actions instead of only returning text.
Cached: model files are already downloaded on this machine.
Use --variants for every hardware-specific variant, provider, size, and cache state.
```

```foundry model list --device npu
╭────────────┬──────┬──────┬────────┬───────┬────────╮
│ Model Name │ Type │ Size │ Device │ Tools │ Cached │
╰────────────┴──────┴──────┴────────┴───────┴────────╯
Type: Chat / Multimodal / Speech / Embedding = model task.
Device: NPU / GPU / CPU = default hardware target selected for this machine.
Tools: supports tool calling; the model can request API calls or app actions instead of only returning text.
Cached: model files are already downloaded on this machine.
Use --variants for every hardware-specific variant, provider, size, and cache state.
```

### Urgency

Low urgency, but was amusing that 0.10.3 fixed my Surface Laptop 7 with Qualcomm but broke my HP Elitebook

### Platform and architecture

Windows X64

### OS Version

Windows 11 Enterprise 25H2

### Installation type

Released package/binary

### Foundry Local version

0.10.3

### API or surface area

CLI

### Hardware acceleration/backend

WinML EPs

### Backend/runtime version

_No response_