Tag
Orthrus is a dual-architecture framework that combines autoregressive LLMs with diffusion models for fast parallel token generation while maintaining exact inference fidelity via shared KV caches and consensus mechanisms, achieving up to 7.8x speedup.