Real-Time Optimization Details for MoE Embodied Video Streaming

omarsar0 · x · 2026-07-11

A reply adds system and architectural details: to achieve real-time performance, the system uses dual-expert distillation, FP8 TensorRT, paged KV-cache, and runtime cleanup, reducing single-chunk latency from 927ms to 142ms for up to a 6.5x end-to-end improvement. Structurally, it employs a sparse MoE video stream with 128 experts, top-8 routing, and a shared expert, totaling about 13B parameters with roughly 1.9B activated per token, thereby balancing complex visual dynamics with high-frequency control.

Related event: LingBot-VA/VLA 2.0 Released: Native Embodied Foundation Model(24 posts)→

Original post →

More from Embodied

Embodied channel →