Apple · open weights · text to video

FastFrames: Apple's open video model, and what it takes to run

Apple published FastFrames as open weights in November 2026: a text-to-video model in two sizes, FastFrames-L (14B) and FastFrames-S (1.3B). This site answers the question people actually ask next — will it run on my hardware, and where can I try it today.

The three questions everyone asks about FastFrames

  1. Which size should I run?

    FastFrames-S (1.3B) is the one to start with: it downloads fast, fits on modest cards and is quick enough to iterate on a prompt ten times. FastFrames-L (14B) is the quality ceiling, and it is the one that forces you to think about quantisation, offloading and batch size.

  2. Will it fit on my card?

    Do the arithmetic before you download anything. In FP16 the weights alone are roughly parameter count × 2 bytes: about 2.6 GB for the 1.3B model and about 28 GB for the 14B one. A 24 GB card can hold the small model with room for the encoder and activations; the 14B model needs quantisation or CPU offloading to fit there.

  3. Can I try it without a GPU?

    Yes, for now that is what the community Spaces are for — they run on Hugging Face's ZeroGPU hardware (A10G-class). They queue, they rate-limit, and they are somebody else's server, but they are live today. Our where to try page tracks which ones are up.

FastFrames-L vs FastFrames-S

Weights-only memory is simple arithmetic, not a measured figure. Total memory use depends on resolution, frame count, precision and offloading, so treat the last column as a starting point and check the model card before you plan a machine.

ModelParametersWeights in FP16Best forRealistic home for it
FastFrames-L14B~28 GBFinal renders, best motion quality40 GB+ GPU, or 24 GB with quantisation and offloading
FastFrames-S1.3B~2.6 GBPrompt iteration, fast previews12–16 GB consumer GPU

What FastFrames actually is

The release is small and easy to describe precisely: two checkpoints, one task, no product wrapper. apple/FastFrames-L carries 14B parameters and apple/FastFrames-S carries 1.3B; both are tagged text-to-video on Hugging Face. There is no subscription, no credit system and no queue of your own to manage — you either run it or you use someone else's Space.

Two sizes, two different jobs

The small checkpoint is the one to reach for first. At 1.3B parameters it downloads in minutes and leaves room on a consumer card for the text encoder and the decoder, which means you can change one word of a prompt and generate again without planning your afternoon around it. The large checkpoint is for the final pass: 14B parameters buy motion coherence and detail that a 1.3B model cannot hold, and in exchange you spend real time on quantisation, offloading and batch sizing.

That split is the whole decision. Iterate on the small model, render on the large one, and do not try to make the large model behave like the small one on hardware that cannot hold it.

What the release does not include

Three things are missing and each one shapes what you should do next. There is no hosted API on the major inference platforms yet — fal's library, with more than 1,500 models, did not list FastFrames when we checked on 18 November 2026. There is no one-click browser demo run by Apple that we could find; the working demos are community Spaces. And there are no vendor-neutral speed numbers, which is why every memory figure on this site is labelled as arithmetic rather than as a measurement.

How to read the hardware advice here

We do not reprint benchmark tables we have not run. What we do instead is show the arithmetic — parameters times bytes per parameter — and tell you which variable to change when it does not fit. Those numbers are honest and checkable. Anything that depends on resolution, frame count or sampling steps is a range, and the model card on Hugging Face is the source of truth for it.

If your card is under 24 GB, start with FastFrames-S and read the local setup guide. If it is 24 GB or above, read the FastFrames-L page first and decide whether quantisation is acceptable to you before you download 28 GB of weights.

FastFrames FAQ

What is FastFrames?

FastFrames is a text-to-video model Apple published as open weights in November 2026. It ships in two sizes: FastFrames-L with 14B parameters and FastFrames-S with 1.3B, both on Hugging Face under the apple organisation.

Is FastFrames free to use?

The weights are published openly on Hugging Face at no cost to download. The terms of use are on the model card, and this page does not restate them — read the licence there before you use output commercially.

Can I try FastFrames without a GPU?

Yes. Community demos on Hugging Face Spaces run the model on ZeroGPU hardware, so you can generate from a browser. They queue and rate-limit, and they belong to their authors — our where-to-try page lists which ones are up.

Will FastFrames-L run on a 24 GB card?

Not at full precision. The 14B weights alone are about 28 GB in FP16, so on a 24 GB card you need quantisation or CPU offloading, and the clip length and resolution you can afford drop. FastFrames-S at roughly 2.6 GB of weights fits comfortably.

Is there a hosted FastFrames API?

Not on fal.ai as of 18 November 2026 — the model is not in its library. Hosted access usually arrives within days to weeks of a popular open release, so this is worth rechecking rather than assuming.