[R] LLaDA2.1 vs Qwen3 30B A3B: Benchmarking discrete diffusion LLMs against autoregressive MoE models

The leaked LLaDA2.1 paper reveals a discrete diffusion language model that reportedly edges out Qwen3 30B A3B in quality (73.54 vs 73.09 avg) while achieving dramatically higher throughput. In its quantized 'S Mode', it hits 674.3 tokens per second versus Qwen's 240.2 TPS. The model introduces a novel T2T editing mechanism and EBPO RL framework to correct errors during parallel decoding, addressing key inconsistency issues from prior diffusion models.