modeldelta

See what changed inside any model checkpoint.
Weight deltas, SVD structure, and a diagnostic verdict.

Compare models Browse gallery
$ pip install modeldelta

How it works

No GPU. No full model loading. Just math on weight deltas.

Stream tensors

Downloads only safetensors files. Loads one tensor at a time — never both full models in RAM.

Compute deltas

For each weight matrix: Frobenius norm, cosine similarity, sparsity. Memory-optimized in-place subtraction.

SVD analysis

Full singular value decomposition of each delta. Effective rank, top-k concentration, spectral decay.

Diagnose

Automatic classification: surgical, standard, heavy, or extreme. Human-readable conclusions, not just numbers.

Diagnostic profiles

modeldelta classifies fine-tuning intensity into four categories.

SURGICAL ΔW/W < 1.5% — Minimal, targeted changes. Qwen-style.
STANDARD 1.5–5% — Typical SFT. Llama-3.1, Mistral.
HEAVY 5–12% — Aggressive training, LayerNorm may shift.
EXTREME >12% — Full rewriting. Gemma-2-9B territory.

Compare your own models

Choose the method that fits your models.

💻 Local CLI

Any model size. Needs ~6.6 GB RAM for 7B. CPU-only, no GPU required.

modeldelta Qwen/Qwen2.5-3B Qwen/Qwen2.5-3B-Instruct -o report.html

☁ Google Colab

For models >3B. Free CPU runtime. ~18 min per 7B pair.

⚡ Online (≤3B)

Small models compared directly in the browser. No install, no Colab needed.

Open HF Space →