4 models · 3.3 MB each · one voice
Four models, one character at a time.
13 MB of weights, loaded on demand and run on your own GPU.
the vectors they exchanged
a 2-D shadow of the 128-D hidden states
what came out
the merged decision
What it is
Four small language models, each branched from one ancestor and then fine-tuned on a different quarter of Shakespeare. They run at the same time on your GPU and merge their hidden vectors — not their text — into one character.
Why we built it
Because the embedding table is frozen and shared, all four think in the same coordinates, so their vectors can simply be added together. That makes the hand-off visible: which model was certain, what it contributed, and where the merged answer landed.
What it changes
If models can share one basis, a network stops being a single file you retrain whole and becomes parts you can swap, add, or retire one at a time. The merge stays something you can watch rather than something you have to trust.
All four run for every character — this is not a speed trick, and we make no claim that it beats one model of the same total size.