HomeAIAI ModelsApple Releases FastFrames Open Video-Generation Models

Apple Releases FastFrames Open Video-Generation Models

Apple has released FastFrames, a family of open video-generation models aimed at interactive latency, as a research drop rather than a consumer product. The weights, code and paper appeared this week. The speed and quality figures below are the authors’ measurements, as stated in the paper and on the project page, not independent tests.

The FastFrames-S model card on Hugging Face is dated 16 November 2026. A paper, FastFrames: Video Generation at Interactive Latency, was posted to arXiv the next day as 2611.23226, and Apple’s machine-learning research group has published a project page with videos, code and model weights. The release covers two variants: FastFrames-S, with 1.3 billion parameters, and FastFrames-L, with 14 billion. Both are described as autoregressive video diffusion models that generate a frame in two sampling steps. The project materials say the models output 832×480 video, in clips of 81 frames at 16 frames a second.

The paper’s headline timing claims are specific. It says the 1.3 billion parameter model runs at 36 milliseconds a frame, or 28 responses a second, on a single NVIDIA B200 GPU, “while matching the generation quality of a state-of-the-art model that is nearly 8× slower.” It says the 14 billion parameter model runs at 45 milliseconds a frame, or 22 responses a second, on an 8× B200 node, and “yields significantly better visual quality and realism.” The Hugging Face card’s worked example for FastFrames-S lists a total of 35.4 milliseconds, in the same range as the paper’s 36 millisecond figure. The authors also write that “LingBot 1.3B optimized with FastFrames delivers a 16× speedup.” Those comparisons are the paper’s.

The technical argument is about where the time goes. The authors say they profiled the inference pipeline of current autoregressive video diffusion models and found that steps often treated as fixed implementation details, including VAE decoding and key-value cache updates, “account for over half of end-to-end inference latency.” They say they replaced those components with lighter alternatives. FastFrames-L is described as having been distilled from Wan2.1-14B. The paper frames “interactive latency” against human-computer interaction studies that have reported delay-detection thresholds of about 50 to 60 milliseconds, and presents a bar of 50 milliseconds or less, equivalently 20 or more responses a second, as its test for seamless interactivity. That threshold is the authors’ chosen standard, not a product specification from Apple.

The work is credited to Gene Chou, listed with Apple and Cornell University, together with Apple co-authors Guandao Yang, Fujun Luan, Lars Mescheder, Wei Dong, Stephan R. Richter and Vladlen Koltun. The paper says the work was completed during Chou’s internship at Apple. The weights are not an unrestricted open-source grant. FastFrames is released under the Apple Sample Code License, and the project materials say parts derived from LingBot are under a Creative Commons Attribution-NonCommercial-ShareAlike 4.0 licence. That combination allows researchers to inspect and run the models under those terms. It is not a general permission to use the weights in a commercial product.

Licence limits of that kind have become part of the open-video-model story TSN has already been covering. Pika’s H3 Flex weights were opened with a licence that excludes the UK and the US. SenseTime’s Code2Video was released as an open model that generates video from code. On speed, FastH3-Omni’s authors reported large cuts in MiniMax H3 generation times on their own tests. FastFrames belongs in that file of research and open-weight video releases, not in a list of new Apple consumer features.

A research release is not a product, and the speed figures are the paper’s

Nothing in the card, the project page or the paper describes FastFrames as a feature inside an iPhone, Final Cut Pro, or any other shipping Apple product. It is a research family with published weights, and that is the right way to read it. The 36 millisecond and 45 millisecond figures, the “nearly 8× slower” quality comparison, the 16× LingBot speed-up and the 50 millisecond interactivity bar are all the authors’ own claims, measured on B200 hardware they specify. Until someone outside the project runs the same tests, those numbers should be quoted as the paper’s results, not as a settled ranking of video models.

The licences point in the same direction. Sample-code terms and a non-commercial ShareAlike clause on LingBot-derived parts are a deliberate restriction, not a footnote. Anyone hoping to build a product on top of FastFrames will have to read those licences as carefully as the latency tables. The next facts to watch are therefore narrower than a product launch: whether independent groups can reproduce the interactive-latency claims, whether Apple later ships a product that actually uses this work, and whether the licence terms are widened or left as a research grant.

Sources

Related stories

Pika Opens H3 Flex Video Weights, but Licence Excludes UK and US is the closest recent TSN account of an open video-model release whose licence terms matter as much as the weights, and is the next read.

Follow @tsnmediahub on X for breaking AI news as it happens.

Follow on X

Share this story

Latest stories

More in this category

Latest stories

Explore the TSN Atlas: Every topic we cover, mapped. Click a bubble to dive in.