# The candidate: the hybrid with HC, a value residual and a 4-tap KV conv.
# The dense FFN stays in the first layer. Shown at d=1024 (16 layers).
use m2026_hybrid
d_model = 1024

global = fr_global(kv_conv_size=4)
candidate = embed
            fr_layer(hy_delta, fr_dense) fr_layer(hy_delta)^2 fr_layer(global(v=@v1))
            (fr_layer(hy_delta)^3 fr_layer(global(v=$v1)))^(layers // 4 - 1)
            fr_norm @h fr_out hy_mtp
model = candidate
residual = hc(expansion=4)