ReplaySSM Hits Production in GLM-5.3-Flash and Kimi-K3 With Up to 2.1x Faster Decoding

Tri Dao and NVIDIA announced that ReplaySSM is now in production at GLM-5.3-Flash and Kimi-K3 and integrated into TensorRT-LLM, delivering up to 2.1x faster decoding; Albert Gu noted alternative architectures still have far more room for optimization than transformers.

2026-10-13 ~ 2026-10-13 · 2 related posts

• ReplaySSM ships in GLM-5.3-Flash and Kimi-K3, speeding decoding up to 2.1x — tri_dao · 2026-10-13

1 near-duplicate retellings: SonglinYang4