# FastFrames — Apple's video models generate 480p frames in under 50 ms

> Apple speeds up autoregressive video diffusion to interactive latency by fixing the decoder and KV-cache bottlenecks.

- Org: Apple
- Date: 2026-11-17
- Category: model, paper
- Source: https://huggingface.co/apple/FastFrames-L
- Canonical page: https://ai-tldr.dev/releases/apple-fastframes/

## In short

Apple's FastFrames are 1.3B and 14B autoregressive video diffusion models that reach interactive latency at 480p: 36 ms per frame on one B200 for the small model, 45 ms on an 8x B200 node for the 14B.

## What it is

FastFrames is a pair of text-to-video models from Apple researchers: FastFrames-S (1.3B) and FastFrames-L (14B, fine-tuned from Wan2.1-T2V-14B). Both generate 480p video one chunk at a time fast enough to react to input, and FastFrames-L sits in the Hugging Face trending top 10.

## How it works

The paper profiled existing streaming video pipelines and found VAE decoding and KV cache updates took over half of end-to-end latency. FastFrames swaps them for a KV cache refinement step and a lightweight pixel decoder, plus systems work that speeds up diffusion sampling. The authors report one generation step falling from 271.8 ms to 35.8 ms.

## Why it matters

The 1.3B model matches the quality of a model nearly 8x slower, and the project page shows the method speeding up LingBot camera-controlled generation by 16x, so it applies to interactive world models, not only clips. Licensing is restrictive: Apple's research license on the weights, Apple Sample Code License on the code, and CC BY-NC-SA 4.0 for the LingBot parts.

## Who it's for

researchers building real-time video and world models

## Try it

https://github.com/apple-aiml-research/ml-fastframes

## Links

- [Model Card](https://huggingface.co/apple/FastFrames-L)
- [arXiv Paper](https://arxiv.org/abs/2611.23226)
- [Code](https://github.com/apple-aiml-research/ml-fastframes)
- [Project Page](https://apple-aiml-research.github.io/ml-fastframes/)
