# Cambrian-BEV speeds up video reasoning with bird's-eye-view tokens

> A new tokenization method compresses 128 video frames into 729 spatial tokens, cutting LLM prefill time 2.9× and memory 3.6×.

Oossa · 2026-10-19 · https://oossa.com/en/cambrian-bev-speeds-up-video-reasoning-with-bird-s-eye-view-tokens

Researchers introduced Cambrian-BEV, a way to map visual patches from video into a shared bird's-eye-view grid. It uses depth and camera pose to merge observations of the same place, regardless of viewpoint or time. The method keeps the same model architecture but replaces frame tokens with 27×27 BEV tokens. On the VSI-Bench test it scored 63.48, 8.5 points higher than a frame‑token baseline while using far fewer tokens.

## The facts

- 128 frames → 729 BEV tokens (27×27 grid)
- Prefill time reduced 2.9×; peak memory reduced 3.6×

## Why it matters

It lets developers build faster, lighter video‑based AI assistants that can reason about space more reliably.

## Sources & references

1. [Thinking from Above: Efficient Spatial Reasoning with Bird's-Eye-View Visual Tokens](https://arxiv.org/abs/2610.22700) – arXiv cs.CV, 2026-10-19

Last updated: 2026-10-19