Model Explorer
Architecture mode (the default view) shows the complete tensor inventory of a model in a 3D point cloud. Each point is one tensor. The layout encodes layer depth on the Y-axis and tensor kind on the X-axis.
Tensor list
The right panel lists all tensors from GET /architecture — 290 tensors for the reference model. For each tensor, the panel shows:
- Full tensor name (e.g.,
model.layers.0.self_attn.q_proj.weight) - Shape (e.g.,
[896, 896]) - Dtype (
float32) - Parameter count (e.g.,
803,712)
Click any tensor in the list to focus the camera on the corresponding point in the 3D view.
GGUF quantization diff
Drag two .gguf files onto the canvas to enter diff mode. The view shows the absolute difference in dequantized values between the two files, rendered as a histogram per tensor. This lets you compare quantization quality between different quant types (e.g., Q4_K vs Q8_0).
Both files must be the same model architecture for the diff to be meaningful. The diff is computed client-side from real dequantized tensor values — no server required.
Model info pane
The model info section in the right panel shows metadata from GET /model-info:
| Field | Value (reference model) |
|---|---|
| Model | Qwen/Qwen2.5-0.5B-Instruct |
| Device | mps / cuda / cpu |
| Layers | 24 |
| Attention heads | 14 Q / 2 KV |
| Hidden size | 896 |
| Max tokens | 40 |