# The final model: the candidate with its dense FFN moved to the last layer.
# Shown at d=1024 (16 layers), matching the architecture in x_post.md.
use m2026_hybrid
d_model = 1024

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