Reka AI's Rho-1 architecture uses two expert weight streams within each transformer block—an understanding stream for language and visual parsing, and a generation stream for denoising latents—which share attention and a single KV cache