webWeightSpace

Program a language model by editing its weights, not its prompt. A few lines of code compile into WebGPU compute passes that rewrite the parameters: merges, ablations, noise, concept injection. Then WebLLM loads the edited model in the same tab. Nothing leaves your machine.

checking WebGPU…

1Model

2Weight program

Language reference

<selector> <op> [args], one rule per line. Rules apply in order.

Selectors: * · embed · layers.10-13 · layers.*.mlp · layers.*.attn.o_proj · layers.0-5.*norm

Ops: mix a w += a·(donor − base) · scale k · zero · noise s · skip (zero o_proj + down_proj) · zero-heads 0-6 · inject word c (solve value-bias deltas so the layer writes c·unit(E[word]) into the residual stream)

3Prompts

greedy decoding, identical for both models

Weight map

Outlined cells are the tensors your program touches.

Original vs. edited weights

Run a program to see both models answer the same prompts.

Log