TOKEN MIXER

The part of a layer that moves information between positions — the job attention does, and the job a recurrence does instead. A model need not use the same one in every layer.

Naming the job separately from the mechanism is what lets a hybrid stack be described at all: Qwen3.8 runs a Gated DeltaNet recurrence in forty-eight layers and gated attention in sixteen, in a repeating pattern of three then one. The graph declares each layer's operator explicitly and the reader requires it, so the programme is read rather than inferred from which tensors a layer happens to ship — and because the two programmes compute different things with different operands, they need not carry the same precision. This is the concept that makes a precision map necessary rather than merely possible.

IN THE GRAPH