Apple FastFrames-L: 14B Open Video Model, 480p

The short answer
Apple posted FastFrames-L (14B) and FastFrames-S (1.3B) on Hugging Face on 2026-11-16, and @_akhaliq flagged the release on 2026-11-17. Both are two-step autoregressive video diffusion models that output 832 x 480 video at 16 fps, and FastFrames-S is profiled at 35.4 ms total on one GPU. Weights and code are in Apple's repositories under the apple-amlr license tag; no inference provider deploys the model.
On this page (10 sections)
What is Apple FastFrames-L?
FastFrames-L is a 14B open text-to-video model from Apple, released with a smaller 1.3B sibling, FastFrames-S. Both are autoregressive, two-step video diffusion models that generate 480p video at interactive latency, according to the Hugging Face repository.
@_akhaliq wrote on 2026-11-17 that "Apple just released FastFrames," linking both repositories. The FastFrames-L repository itself was listed on 2026-11-16 and is tagged text-to-video.
The accompanying paper is titled "FastFrames: Video Generation at Interactive Latency." The authors are Gene Chou, Guandao Yang, Fujun Luan, Lars Mescheder, Wei Dong, Stephan R. Richter and Vladlen Koltun.
The "interactive latency" framing is the point of the release. Most video models are judged on clip quality and take seconds to minutes per clip. Apple is pitching a model fast enough to respond to input as it is used, and the repository includes an interactive demo to show it.
FastFrames release date and availability
FastFrames-L and FastFrames-S appeared on Hugging Face on 2026-11-16 as downloadable weights with an official code repository. The model page states that no inference provider deploys it, so running it means setting it up yourself.
The requirements are Linux x86-64, an NVIDIA GPU and a CUDA 12.8-compatible driver. Installation uses uv sync, followed by downloading FastFrames and third-party weights using the checkpoint guide.
Apple's latency script for the large model is run as benchmark.py ff-l --gpus 8, and the LingBot demo has both a 1 GPU and an 8 GPU launch command. Apple has not published a minimum hardware spec for FastFrames-L, and no latency figure for the 14B model appears on the page.
FastFrames-L key features and specs
FastFrames-L distills Wan2.1-T2V-14B into a two-step causal model following the Causal Forcing++ recipe, then trains a pixel decoder and a KV refinement network on top. It outputs 832 x 480 video, 81 frames at 16 fps, which is roughly 5 seconds per clip.
Stopwatch resting next to a strip of film frames on a dark desk
| Spec | FastFrames-L | FastFrames-S |
|---|---|---|
| Parameters | 14B | 1.3B |
| Base model | Wan2.1-T2V-14B | Wan2.1-T2V-1.3B |
| Type | Autoregressive, two-step video diffusion | Autoregressive, two-step video diffusion |
| Output | 832 x 480, 81 frames, 16 fps | 832 x 480, 81 frames, 16 fps |
| Measured latency | Profiled with 8 GPUs, no figure given | 35.4 ms total, 1 GPU, 480p |
| Audio | Not announced | Not announced |
| Image-to-video | Not announced | Not announced |
| Hosted API | None listed | None listed |
The FastFrames-S breakdown, with a full KV cache on one GPU, is two DiT evaluations at 16.1 ms and 16.2 ms, 2.2 ms for KV refinement and 0.9 ms for the pixel decoder. That adds up to 35.4 ms, against 36 ms in the paper. The page does not say what unit of output that latency covers, so read it as a speed signal until you have read the paper.
Apple released every training checkpoint for the large model: a causal teacher trained on latents generated by Wan2.1-14B, the consistency-distilled model, the DMD-distilled model, and the final generator with its pixel decoder and KV refinement network. That makes the pipeline reproducible, not just usable.
Apple also released an optimized version of LingBot 1.3B that it says is 16x faster than the official model and allows fluid camera-conditioned control. The repository says it builds on code from Wan2.1, Self-Forcing and Causal Forcing++.
FastFrames license, pricing and commercial use
FastFrames-L has no listed price. The weights are in a public Hugging Face repository tagged license:apple-amlr, and the page says all FastFrames code is under the Apple Sample Code License.
One part is clearly restricted. LingBot-World v2 code and weights follow their upstream CC BY-NC-SA 4.0 license, which is non-commercial. Anything built on the LingBot demo should be treated that way.
The terms for the FastFrames weights themselves are not spelled out in the page text. Until you have read the license file, assume nothing about commercial use of outputs, and have legal review it before FastFrames clips go into paid campaigns.
How to use FastFrames for ad creative
At 480p and 16 fps, FastFrames is a drafting model, not a delivery model. Its draw is how quickly it generates, which lets a creative strategist try many hook concepts, camera moves and scene ideas before spending credits on a higher-resolution model.
Creative team reviewing rough video drafts on a wall of monitors
Where it could fit in a workflow:
- Storyboarding hooks and first-three-second concepts, then re-rendering the winners in a higher-resolution model. The hook rate and hold rate benchmarks show what is worth testing.
- Rough b-roll exploration for product scenes and lifestyle settings.
- Camera-controlled previews through the LingBot demo, for non-commercial exploration only given its license.
Several things that D2C teams need are not in the release. The repository lists text-to-video only, with no image-to-video mode to animate a product shot. There is no audio, and output is capped at 81 frames, about 5 seconds. Product fidelity, text rendering and brand consistency are not addressed on the page.
There is also a practical barrier. Setup means a Linux machine with NVIDIA hardware, which most performance marketing teams do not have on hand. For static ad concepts, the guide to making static ads with AI covers a workflow that needs no local GPU. For video format ideas, see Meta ad creative ideas.
What to watch next
The first thing to watch is the VBench numbers. The repository says Table 1's Quality, Semantic and Total scores can be reproduced with the evaluation guide, which gives outside testers a way to check Apple's claims.
Also open: a latency figure for FastFrames-L, a statement of the weight license terms, and whether Apple adds image-to-video or audio. Apple has given no word on a consumer product or pricing.
Hosted availability is the other open item. The Hugging Face page invites readers to ask for inference provider support, and for now none deploys the model.
What people are posting
Apple just released FastFrames Video Generation at Interactive Latency FastFrames-L (14B): https://huggingface.co/apple/FastFrames-L FastFrames-S (1.3B): https://huggingface.co/apple/FastFrames-S https://t.co/5qnFL9PAEd
Open model repository apple/FastFrames-L, trending on Hugging Face. Tags: safetensors, text-to-video, base_model:Wan-AI/Wan2.1-T2V-14B, base_model:finetune:Wan-AI/Wan2.1-T2V-14B, license:apple-amlr
Key takeaways
- FastFrames-L is a 14B text-to-video model finetuned from Wan2.1-T2V-14B, with a 1.3B sibling called FastFrames-S.
- Both models are autoregressive, two-step video diffusion models that generate 480p video.
- The default script outputs 832 x 480 video, 81 frames at 16 fps, about 5 seconds.
- FastFrames-S totals 35.4 ms on one GPU at 480p with a full KV cache (paper: 36 ms).
- The repository carries the apple-amlr license tag, the code is under the Apple Sample Code License, and no inference provider hosts the model.
- Apple also released a LingBot 1.3B version it says is 16x faster than the official model.
Compare 200+ AI video and image models side by side
Run Kling, Veo, Sora, Seedance, Flux, Ideogram and Recraft in one Playground, or turn any product URL into ready-to-launch ads.
Frequently asked
What is Apple FastFrames-L?
FastFrames-L is a 14B open text-to-video model from Apple, distilled from Wan2.1-T2V-14B into a two-step causal model. It generates 480p video at what Apple calls interactive latency.
When was FastFrames released?
The Hugging Face repository for FastFrames-L is dated 2026-11-16, and @_akhaliq posted about the release on 2026-11-17. Both FastFrames-L and FastFrames-S were published together.
Can I use FastFrames-L for commercial ads?
The license terms for the weights are not spelled out on the repository page. It carries the apple-amlr license tag, and the code is under the Apple Sample Code License. Read the license file before using outputs in paid ads.
How do I run FastFrames?
It needs Linux x86-64, an NVIDIA GPU and a CUDA 12.8-compatible driver. Install with uv sync, download the checkpoints, then run scripts/generate.py with --model ff-l. The Hugging Face page says no inference provider deploys it.
What resolution does FastFrames output?
The generation script outputs 832 x 480 video with 81 frames at 16 fps. Apple describes both models as 480p, so it suits drafting more than final delivery.
Sources
- apple/FastFrames-L · Hugging Facehuggingface.co
- apple/FastFrames-S · Hugging Facehuggingface.co
This story is still unconfirmed: it is based on what people are posting, and the company behind it has not announced it yet. We will update this page as it develops.
More AI model news & creative guides

Recraft Studio Font in Prompt: Attach Any Typeface
Recraft Studio now lets you attach any typeface, from its Font library or your own brand font, to a prompt. Get the ad workflow breakdown and access details.

Krea Video Swap: Swap Characters, Products, Backgrounds
Krea launched Video Swap on Nov 17, 2026 to change characters, products and backgrounds in video. Get the spec status, ad workflow breakdown and access details.

Muse Code 3D: Native GLB Generation, Limits & Access
Meta added native text and image to GLB 3D generation to Muse Code on Nov 20, 2026. Get the spec table, ad workflow breakdown, plan limits and access details.
Best AI Ad Generators for D2C Brands in 2026
Ten AI ad generators compared for D2C and e-commerce brands: what each makes, real starting prices, free options, and which suits Indian brands paying in rupees.
Break Even ROAS for COD Orders: How RTO Changes the Math
How to work out break even ROAS when a share of COD orders comes back as RTO, with worked rupee examples, a prepaid and COD mix, and why Ads Manager ROAS overstates COD.
What Indian D2C Brands Run on Meta: 2,790 Ads Analyzed
We studied 2,790 Facebook and Instagram ads from 14 Indian brands and their 747 longest-running ads: formats, angles, offers, languages and CTAs that last.
Your move
See your product as a winning ad
Turn a product URL into static ads and UGC-style videos, rebuild the competitor ads that already work, and read your Meta data to know what to make next.
AdGraam News DeskAI video and image news
The AdGraam News Desk (part of AdGraam, also searched as Adgram or Adgram AI) tracks AI image and video model launches, leaks and rumors from labs, leakers and creators, and checks each claim against its sources. Unconfirmed stories say so, and each page is updated as its story develops.