Weights Visualizer

Model — Params — Tensors — Experts —

Reference architectures

25 major real models · how each one actually computes

Architecture

parameters per layer · in forward order

attention experts / MoE embed / head norm / other

Tensor explorer

tensor	shape	dtype	params	mean	std	≈0	‖w‖₂

Weight matrix

< 00> 0

Value distribution

Analysis lab

Calibration per-channel

original dequantised error

Outlier threshold 4σ

Magnitude tail · log-count per |w| decile band

Largest magnitudes · index

Sub-grid Rank 16

σ_i / σ₁ · log₁₀ energy retained chosen rank

A B

metric	A	B	Δ

Checkpoint diff

load a second file to compare Changed if |Δ| >

Change by layer group · forward order

Most-changed tensors · relative &lVert;Δ&rVert;_F

tensor	shape	mean \|Δ\|	rel &lVert;Δ&rVert;_F	cos(A,B)	% changed

Load a fine-tune next to its base model — matched by tensor name — to see exactly which weights moved and by how much. Nothing is uploaded; both files are read in this tab.

Depth sweep

one component, every layer Component Statistic

Token routing & expert flow

simulated forward pass Layer Top-k 2 Tokens 16

Each token is routed to its top-k experts by a softmax gate; the fill inside each expert shows its cumulative load. Cold experts are the pruning candidates — carry the same model into the Experts Pruner.

Reading checkpoint

Drop weights to load

.safetensors, or a torch .pt / .pth · nothing leaves your machine