ZERO.4 · A 4,852,992-PARAMETER TRANSFORMER
Begin with a fragment.Let the machine dream onward.
Promoted after a preregistered three-seed faculty and replay evaluation. Written in C. Running locally in this browser—your words are never sent anywhere.
∅
This is a channel-native literary model. It compresses each exchange into a learned, lossy memory and can recall a relevant older compressed episode as a holographic echo.
The zero opened its eyes…
To see a world…
What spirit walks…
×
MODEL NOTE
A real, very small language model.
ZERO.4 is a decoder-only character transformer initialized from frozen ZERO.3, then trained to route checked quantity operations while preserving its six-source historical replay budget. All three preregistered seeds passed the public and promotion gates; the prospectively selected seed-2 checkpoint is deployed here. Its matrix rows use 8-bit values with floating-point row scales, while normalization gains remain floating point.
The C engine performs embeddings, RMS normalization, rotary attention, GELU feed-forward layers, and sampling inside WebAssembly. It uses no server and no machine-learning runtime.
This page exposes the promoted generative core and its channel memory modes. The measured quantity capability uses a separate controller to bind arguments and a deterministic kernel to compute arithmetic; neither is simulated by free-form browser sampling. You can compare a recent transcript window, recurrent compression, flat holographic recall, and experimental partitioned recall without changing the 4.85M transformer. It still cannot retrieve outside facts or match a contemporary large language model. Oddness is part of its honest scale.
CONTEXT 512 characters
LAYERS 6
DIMENSION 256
TRAINING loading…
READ THE C SOURCE ↗