Hybrid AI video models just got a lot faster — without the usual accuracy hit.
Researchers studying vision-language models (VLMs) designed for long video found that token reduction — cutting down the number of data chunks a model processes — works differently in hybrid Mamba-Transformer architectures than in standard Transformers. In pure Transformer models, dropped tokens are simply gone. In hybrid models, Mamba layers keep a running memory state that holds onto information from those dropped tokens, so reduction acts more like compression. The team built a probing method to measure how much discarded information survives, then used those findings to design a new reduction framework that applies cuts progressively, starting light and going heavier in later layers where it's safer. On long-video benchmarks, the approach hit 3.8x to 4.2x prefilling speedups at a 25% token budget while staying close to baseline accuracy.
This matters because video is one of the heaviest workloads in AI right now, and the gap between what models can do and what they can do cheaply is a real bottleneck for deployment. Most token-reduction research targets Transformer-only architectures, so this work opens a practical efficiency path for hybrid models that have been gaining traction but lacked tailored optimization methods. Unusually, the hybrid models here actually improved in both speed and accuracy under aggressive reduction — the opposite of the typical trade-off.
Hybrid architectures have been pitched as a middle ground between Transformers and state-space models for a while; it's useful to see efficiency research finally catching up to that claim.