3D was chosen deliberately. X and Y are the two matrix axes, so a tensor with shape [out,in] is drawn as an in × out rectangle. Z is execution depth. This keeps every matrix area exact while giving the 896 experts room to exist in parallel.
The physical atlas rotates some rectangles by 90° only to pack them tightly; area remains exact. At the default world scale, every scalar is a 2.5 mm × 2.5 mm tile. The unit selector can deliberately rescale that physical interpretation.
Runtime caches are overlays in both views. One cache scalar remains one square scene unit. MLA key/value caches are split into 32,768-token pages and folded through depth; KDA preserves all 96 recurrent head states plus q/k/v short-convolution history in exact-area aggregate sheets.
Router choices are input-dependent. Bright top-16 lines use a deterministic illustrative set of experts for the active layer; the faint fan-out shows all 896 candidates.
Controls
Left drag: orbit. Right drag or Shift+drag: pan. Wheel/pinch: percentage dolly toward the exact surface under the pointer. A hovered/selected surface becomes a hard 12 cm camera barrier.
Flight mode: drag to look, W/A/S/D to move, E/Q to change altitude, and Shift to sprint. Touch devices get independent move and look joysticks plus altitude and boost controls. Swept collision against learned matrices and visible cache sheets provides wall sliding at a 12 cm standoff. The ground does not clamp flight altitude.
Click a matrix to select it. Double-click focuses it.
F fit all, L active layer, S selected, G toggle flight/orbit, Space play/pause outside walking mode.
Every matrix and cache sheet carries procedural screen-space text. At close range the renderer shows the literal scalar grid plus mock digits. Once scalars go sub-pixel, it transitions to a stochastic footprint filter so zoomed-out values read as tinted white noise instead of moire stripes; the full glyphs continuously resolve again as you approach. These are visual placeholders, not checkpoint values.
Selecting a tensor highlights its downstream matrices and related operation nodes; the inspector lists clickable connections.