The family tree of open-weight LLMs, reconstructed from the weights alone. No model cards were read. Every edge below is statistical evidence computed from ~100–300 MB of sampled weights per model — σ-curves, seeded weight samples, norm vectors, and singular-value spectra.
specimen map
Each dot is a model on the Hugging Face Hub; size tracks parameter count. A solid line means the calibrated probability of shared lineage is ≥ 0.90 — thicker when the pair's parentage is corroborated by the publisher's own documentation. Magenta edges are decomposed merges: the fitted mixture weight on each edge matches the merge recipe published on the model card (recovered from weights alone, then checked against it). Drag to orbit, scroll to zoom, hover to inspect, click to pin.
pairwise evidence
Calibrated P(shared lineage) for every depth-compatible pair, ordered by family. The bright blocks are families rediscovered from weights; the dark off-diagonal sea is why the false-positive rate is zero. Hover any cell for the pair behind it.
merge forensics
A merged model's weights are a (near-)linear mixture of its
parents, so the mixture can be refit from sampled weights. For these two real
merges the fitted weights match the mergekit recipe published on the model
card — the recipe was recovered blind, then checked against the card.
The magenta edges in the constellation above carry the same weights;
reproduce with modeldna decompose or
benchmarks/merge_decompose_bench.py.
ground truth
Real Hub models whose parentage is documented by the publishing org, judged blind against all candidate parents by the actual verdict engine. Wrong-parent and cross-family comparisons serve as hard negatives.