Skip to content

A computer, compiled into the weights of a language model. An exact CPU inside Qwen3.5-4B, with Gated DeltaNet state as RAM.

334 B
Raw Blame
"""CHIP-8 inside a stock Qwen3.5 checkpoint: bf16-exact circuits built from Gated DeltaNet and SwiGLU."""

try:  # the chip is exact only without bf16 split-K reductions, which cuBLAS may pick for some shapes
    import torch

    torch.backends.cuda.matmul.allow_bf16_reduced_precision_reduction = False
except ImportError:
    pass