Apple’s FastFrames-L makes a 14B video model two-step fast
Apple has published FastFrames-L (14B) and FastFrames-S (1.3B) on Hugging Face, two-step autoregressive video models built for interactive latency at 480p.

Apple has published FastFrames-L on Hugging Face, a 14-billion-parameter video model that generates 480p clips in just two diffusion steps. A smaller 1.3B version, FastFrames-S, came out alongside it, and Apple lists its measured latency at 35.4 ms.
Two models, one interactive-latency recipe
According to the model page, both models are autoregressive and use two-step video diffusion. They generate 832 x 480 video, 81 frames at 16 fps, which works out to a little over five seconds of footage. The code is meant for Linux x86-64 machines with an NVIDIA GPU and a CUDA 12.8-compatible driver.

FastFrames-L is built on Wan-AI’s Wan2.1-14B text-to-video model. Apple says it first distilled that model into a two-step causal model following the Causal Forcing++ recipe, then trained a pixel decoder and a KV refinement network around it. FastFrames-S keeps a frozen two-step, frame-wise Causal Forcing++ generator and trains the same surrounding networks.
Apple says it released every stage of the pipeline: a causal teacher trained on latents generated by Wan2.1-14B, a consistency-distilled model, a DMD-distilled model, and the final generator with its decoder and refinement network. That is more than most video releases offer, since researchers can start from intermediate checkpoints rather than only the finished model.
The release also includes an optimized version of LingBot 1.3B that Apple says is 16 times faster than the official model and allows camera-conditioned control. A browser demo launches locally at http://localhost:8000 through apps/lingbot/demo.py, and text-to-video generation runs through scripts/generate.py with --model ff-l for the large model.
The repository’s latency table covers FastFrames-S on one GPU at 480p with a full KV cache. The two transformer evaluations take 16.1 ms and 16.2 ms, KV refinement 2.2 ms and the pixel decoder 0.9 ms, for 35.4 ms in total. The paper reports 36 ms. For FastFrames-L, the benchmark command uses --gpus 8, and the page gives no single-GPU figure or memory requirement.
How to run it and what the license allows
The weights are downloadable now through the checkpoint guide in the repository. Apple’s own FastFrames code is under the Apple Sample Code License. The LingBot-World v2 code and weights keep their upstream CC BY-NC-SA 4.0 license, which restricts commercial use, so the LingBot demo is a research tool rather than something to build a product on. The page does not say whether the FastFrames weights themselves carry commercial terms, and readers should check the license files before any commercial plan.
Hugging Face lists no download counts yet, and it shows one Space using the model. The page names Chou, Yang, Luan, Mescheder, Dong, Richter and Koltun as authors of the paper, which is cited as a 2026 arXiv preprint. The model page links an evaluation guide for reproducing the VBench scores in Table 1, but the scores themselves are not printed there, and the paper was still listed as coming soon.
Numbers in context: From minutes to milliseconds
The base model gives a sense of the gap Apple is claiming to close. The Wan2.1 model card says the 1.3B version needs about 8.19 GB of VRAM and generates a five-second 480p video on an RTX 4090 in roughly four minutes, without optimizations such as quantization. The same card lists the 14B model under the Apache 2.0 license and supports both 480p and 720p output. FastFrames-L starts from that 14B model but is limited to 480p in this release and, by the repository’s examples, is meant to run across several GPUs.
Set against four minutes for a clip, 35.4 ms per measured pass is a different regime. The page does not say exactly what one pass covers. Because the generator works frame by frame, it is plausible that the figure is per frame, but the source does not confirm that, and the comparison is not like for like: the Wan2.1 number is for a different model, a different GPU and an unoptimized pipeline. If the figure does refer to a frame, 16 fps playback would need about 62.5 ms per frame, which would leave headroom on the S model. That is an inference, not an Apple claim.
The 14B model is the harder case. Apple’s page does not give a latency table for FastFrames-L, only an eight-GPU benchmark command, so whether the large model reaches interactive speed on a single card is not established. The quality side is also unverified: the VBench Quality, Semantic and Total scores are Apple’s own and have not been independently reproduced in anything the source shows.
For readers, the practical change is the checkpoint set. Anyone with suitable NVIDIA hardware can now study how a large video diffusion model is compressed to two steps, and test camera-controlled generation with the LingBot demo. Whether the approach holds up on longer clips or at 720p is not covered by the source.
Sign Up for the Daily AI Pulse
One email a day. All the stories that matter.
By signing up, you agree to our Terms of Use and acknowledge the data practices in our Privacy Policy. You may unsubscribe at any time.
Alex CarterAI News & Big Tech CorrespondentAlex Carter writes for Aiholics, keeping readers updated on the fast-paced world of AI and Big Tech.



