Recent releases of diffusion large language models, or dLLMs, such as Inception Labs’ Mercury 2 in February 2026 and Google DeepMind’s DiffusionGemma in June 2026, have demonstrated strong inference speed advantages through parallel denoising rather than sequential autoregression. However, these models remain competitive only with mid-tier autoregressive systems and have not closed the capability gap with frontier transformer-based leaders on benchmarks or real-world tasks. Sustained scaling of dLLMs to match or exceed top models faces technical hurdles in training stability and data efficiency, while major labs continue investing heavily in proven architectures. This leaves little time before 2027 for a dLLM to claim the top spot, though an unexpected breakthrough in continuous diffusion methods or a major lab pivot could still alter the trajectory.