Weightscope
Weightscope
v0.1 · open source
Get the CLI
X-ray for AI model weights

See what actually changed between models.

Stream any two HuggingFace checkpoints, compare them tensor by tensor, and get a real diff report instead of vague release notes and marketing claims.

Tensor-by-tensor diffs
· LoRA + full fine-tune visibility
· Open-source CLI
Live diff preview
llama-base vs llama-medical-v2
14.2% weights changed
Tensor
Avg delta
% changed
layers.0.attn.q_proj.weight
+0.0081
82%
layers.12.mlp.up_proj.weight
+0.0214
94%
layers.23.mlp.down_proj.weight
+0.0157
88%
embed_tokens.weight
+0.0006
3%
lm_head.weight
+0.0103
77%
Total params
7.0 B
Sparse delta
312 MB
Compression
44.8x
Why it exists

The missing git diff for model weights.

Today, model updates ship as giant blobs with fuzzy changelogs. Weightscope makes fine-tunes inspectable, comparable, and measurable.

Model laundering

Spot rebranded or near-identical uploads before they get passed off as something new.

No quality signal

See whether a model barely changed or was deeply fine-tuned across its tensors.

Wasted bandwidth

Expose redundant weight shipping and open the door to smarter deduplication.

Try it

CLI-first, fast to test.

Install the open-source CLI, point it at two model IDs, and generate a markdown diff report in seconds.

~/projects/weightscope

      
Compare public HF checkpoints

Stream any two model IDs straight from the Hub — no downloads, no setup.

Shareable markdown reports

Output drops into PRs, model cards, or release notes — no extra tooling needed.

LoRA vs full fine-tune scope

See whether 0.1% of weights moved or 90% — no more guessing from a model card.

Real change magnitude

Per-tensor max delta + sparsity tells you if a "fine-tune" is actually a full retrain.

Pricing

Simple pricing, cleaner presentation.

Free
$0

For open-source exploration and public model diffs.

  • Open-source CLI & library
  • Public diffs on HF
  • HuggingFace Space (≤ 2 GB)
  • Commercial use OK
Choose Free
Most popular
Pro
$9 /mo

For private checkpoint comparisons and hosted history.

  • Private diffs on your checkpoints
  • Hosted history & share links
  • No model size limit
  • Email support
Choose Pro
Team
$39 /mo

For collaborative model analysis across an org.

  • Up to 5 seats
  • Shared workspace
  • SSO (Google, Okta)
  • Audit log + Slack support
Choose Team
If a subscription lapses, previously created artifacts stay accessible. You simply fall back to Free-tier limits for new ones.

Diff your first model in 60 seconds.

pip install weightscope  then point it at any two HuggingFace checkpoints.