Apple does not release many video models. When it does, the interesting part is usually the constraint it chose to optimize. With FastFrames, released this week, that constraint is latency, and the result is a pair of models that generate video fast enough to react while a scene is still playing out.
FastFrames-S, at 1.3 billion parameters, generates a frame in 36 milliseconds on a single Nvidia B200 in the authors' measurements, which the paper expresses as 28 responses per second. FastFrames-L, at 14 billion parameters, takes 45 milliseconds per frame on a node of eight B200s and produces better visual quality in the authors' evaluation. Both are autoregressive, two-step diffusion models: each new frame is generated from the frames and inputs that came before it, which is the design the paper uses for camera-conditioned and physics-conditioned interaction. Those per-frame figures are generation measurements, not a playback rate (the output is specified at 16 frames per second) and not a measured end-to-end response time to a new input.
Where the time was actually going
The paper's useful finding is about plumbing, not model size. The authors profiled state-of-the-art autoregressive video models and found that two operations usually treated as fixed implementation details, VAE decoding and KV cache updates, account for over half of end-to-end inference latency in the models they profiled. The denoising work that research usually optimizes was not the main cost.
FastFrames replaces both. A lightweight pixel decoder stands in for the VAE decoder, and a small KV refinement network stands in for full cache updates. FastFrames-L applies the same treatment after distilling Alibaba's Wan2.1-14B into a two-step causal model, and Apple releases the intermediate checkpoints, including the causal teacher, so other researchers can inspect each stage. The bundled LingBot demo applies the same optimizations to Ant Group's 1.3B world model, which Apple says then runs 16 times faster than the official release, with camera control responsive enough to steer live.
Read the hardware line before the latency line
The published numbers come from datacenter hardware. The 36 millisecond figure was measured on a B200, Nvidia's current datacenter GPU, and the 14B model's 45 milliseconds on a node of eight of them. The model card's stated requirements are Linux x86-64 with an Nvidia GPU and a CUDA 12.8-compatible driver; it sets no minimum GPU model. What the release does not publish is any consumer-GPU number, and there is no Mac version. Until someone runs the small model on hardware readers own, the practical speed on a consumer card is unknown.
Resolution is the second trade. Output is 832 × 480, 81 frames at 16 frames per second, roughly five seconds of standard-definition video. That is a sensible target for interactive scenes, prototypes and world-model research. It is not a film pipeline, and the release does not present it as one.
The license is the third. Per the model card's text, FastFrames code ships under the Apple Sample Code License, and the LingBot-World v2 code and weights carry CC BY-NC-SA 4.0, which excludes commercial use. The weights are public and the checkpoints are complete, but calling this an open-source release without qualification would overstate what the terms allow.
What it is actually for
Taken on its own terms, FastFrames is a research release with a transferable lesson: in the autoregressive models the authors profiled, the bottleneck had moved from the generator to the scaffolding around it, and fixing the scaffolding bought most of the speedup. Any team running Wan-class or Causal Forcing-class models can test that claim against their own pipeline, because Apple published the profiling breakdown alongside the models. The claim that needs outside verification is the quality match, which so far rests on the authors' own VBench evaluation.
Sources
- [1] FastFrames paper, "FastFrames: Video Generation at Interactive Latency," arXiv:2611.23226 (submitted November 17, 2026) — read first-hand: the profiling finding, both latency figures and their hardware, the quality comparison, the interaction applications.Read source
- [2] Apple FastFrames-L model card, Hugging Face — read first-hand: the component latency table (35.4 ms total, "paper: 36 ms"), the training and checkpoint details, the output specification, the system requirements, and the license section.Read source