Back to Blog
    AI & AutomationNov 22, 2026

    Apple FastFrames: AI Video at Interactive Speed, Open to Everyone

    Amir

    Amir Farid

    7 min read

    Apple FastFrames: AI Video at Interactive Speed, Open to Everyone

    Apple has released FastFrames, a pair of open-weight AI video models that generate 480p video at interactive speed — about 36 milliseconds per frame, according to Apple. That is fast enough to steer a scene live, changing the camera angle or the action as the video plays, instead of typing a prompt and waiting for a render. The models, FastFrames-S and FastFrames-L, are free to download on Hugging Face, with the code and a research paper to match. One catch: the licence is research-only.

    What Apple FastFrames actually is

    FastFrames is a research release from Apple Machine Learning Research, published with Cornell University on 16–17 November 2026. It comes in two sizes: FastFrames-S at 1.3 billion parameters and FastFrames-L at 14 billion. Both are autoregressive, two-step video diffusion models: they generate video one frame at a time, using only two internal steps per frame.

    Both models produce the same output: 832 x 480 video, 81 frames at 16 frames per second — about five seconds per clip. Apple says the small model reaches about 35.4 milliseconds per frame on a single B200-class GPU — the paper rounds it to 36 milliseconds — while the large model reaches about 45 milliseconds per frame across eight GPUs. Weights are on Hugging Face under the name FastFrames-S, with code and paper alongside.

    What "interactive latency" means in practice

    At 16 frames per second, a frame needs to appear every 62.5 milliseconds to play smoothly. FastFrames-S generates each frame in roughly 36 milliseconds, leaving headroom inside that budget. That is the difference between a tool you prompt and one you drive.

    Apple also released an optimised LingBot-World v2, a 1.3 billion parameter camera-conditioned world model that Apple says runs 16 times faster than the official model, with live keyboard and mouse control of a 6-degrees-of-freedom camera — virtual-production scouting on a single GPU.

    How FastFrames works, in plain terms

    A video production desk with a neon city film frame generating on the monitor and a camera on a gimbal

    Most video diffusion models refine each frame through dozens of small steps. FastFrames gets to two steps by distillation: Apple trained a slow, high-quality "teacher", then taught a smaller "student" to reproduce its output in just two steps, trading a little quality for enormous speed.

    FastFrames-L is distilled from Wan2.1-T2V-14B, an open text-to-video model. FastFrames-S freezes a frame-wise Causal Forcing++ generator, which already generates frames one at a time, and trains the networks around it. Apple says quality matches or beats slower baselines — its own claim, from its own evaluation.

    The catch: research only, and serious hardware

    Read the licence before planning anything around FastFrames. The models are released under the Apple Research Model License, which permits research use only — not commercial use. The code is under the Apple Sample Code License; the LingBot-World v2 release is under CC BY-NC-SA 4.0, also non-commercial. You can download, run and study the weights, but not put them in a product you sell or use them for client marketing videos without a different licence from Apple.

    Then there is the hardware: Linux on x86-64, an NVIDIA GPU, a CUDA 12.8-compatible driver, and the uv package manager. FastFrames-S needs a single B200-class GPU — a data-centre card, not a laptop GPU — and FastFrames-L needs eight. This is a research release for teams with serious GPU access. It is not a production marketing-video tool yet.

    What interactive video changes

    If video can be generated at interactive speed, the workflow around it changes completely:

    • Previsualisation: a director can block out a scene and move the camera live.
    • Product demos: a configurable product shown from any angle, generated on the spot.
    • Games and virtual worlds: a world model responding to keyboard and mouse input — a game you explore rather than a video you watch.
    • Virtual production: scouting a location or rehearsing a camera move inside a generated world before anyone hires a set.
    • Training and simulation: scenarios responding to the trainee — a warehouse, a retail floor, a driving situation — generated live rather than scripted.

    None of these need 4K resolution or minute-long clips. The resolution ceiling matters less than the latency floor — and FastFrames attacks the latency floor directly.

    FastFrames vs the hosted video tools

    FastFrames differs from the hosted tools Coheser has covered recently in one structural way: it is open weights you run yourself, not a hosted API you pay per generation. Google's Gemini Omni 1.2 Flash, covered on 6 November, is a hosted multimodal model with video generation behind an API; Kling Image 4.0, covered on 18 November, is a hosted image model aimed at 8K output. Both are built for production use today. The same question comes up in image work — see our guides to ChatGPT Images 2.6, MiniMax H3 and Ideogram 4.6.

    FastFramesHosted tools
    AccessOpen weightsHosted API, pay per use
    Speed~36 ms per frame, interactiveMinutes per clip
    Output480p, ~5-second clipsUp to 4K–8K, longer clips
    LicenceResearch onlyCommercial use allowed
    HardwareLinux + NVIDIA GPUNone — cloud-based
    Best forResearch, previs, prototypesShipping content now

    A hosted tool is what you use to make a video this week. FastFrames tells you the hosted tools will get faster and more interactive.

    What this means for your business now

    For most businesses, the right response is to watch FastFrames, not adopt it: no commercial use, and a high hardware bar. Three things are worth doing now.

    First, if you commission video, ask your suppliers about interactive video — the tools experimenting with real-time generation today will offer different formats next year. For build costs, see our guide to AI agent development costs in the UK, our custom software development cost guide and our mobile app development cost guide for the UAE.

    Second, think about where live, steerable video would change your product: a configurator rendering a product from any angle, a training simulator responding to the user, a virtual showroom to walk through. These start as a custom software project — test the idea with a small prototype, as our guide to scaling a SaaS MVP without rebuilding explains.

    Third, keep research and production apart. FastFrames is a signal about the future, not a tool for this quarter. Ship with the production tools that exist today — a custom AI agent for customer conversations, automation for repetitive work, or a mobile app in the customer's pocket. If you want help deciding, talk to us.

    Frequently asked questions

    Can I use Apple FastFrames for commercial videos?

    No — not under the current licence. FastFrames is released under the Apple Research Model License, which is research only. You can download, run and study it, but you cannot use it in a commercial product or for client marketing videos. For commercial video today, the hosted tools in our Gemini Omni 1.2 Flash guide and our Kling Image 4.0 guide are the practical route.

    How fast is FastFrames, really?

    Apple reports about 35.4 milliseconds per frame for FastFrames-S on a single B200-class GPU, and about 45 milliseconds per frame for FastFrames-L across eight GPUs. At 16 frames per second, a frame must appear within 62.5 milliseconds — the small model fits inside that budget. These are Apple's figures from its own testing.

    What hardware do I need to run FastFrames?

    The published requirements are Linux on x86-64, an NVIDIA GPU with a CUDA 12.8-compatible driver, and the uv package manager. In practice, FastFrames-S needs a B200-class data-centre GPU; FastFrames-L needs eight. This is not a laptop project; it is aimed at research teams with serious GPU access.

    How does FastFrames compare with Gemini Omni or Kling?

    They solve different problems. Gemini Omni and Kling are hosted, commercial tools producing high-resolution output you can ship today, priced per use. FastFrames is open weights you run yourself, at lower resolution and shorter clips, under a research-only licence — but at interactive speed. Think of it as a preview of where video generation is heading.

    When will interactive AI video be usable in real products?

    There is no announced date, and Apple has not said it will commercialise FastFrames. The realistic path is that the techniques — two-step distillation and causal frame-by-frame generation — reach commercial models and hosted APIs over the next year or two. Businesses that want to be ready should prototype interactive formats now, for example through AI automation and custom software builds.

    #Apple FastFrames#AI video#open-weight AI#video generation