Apple’s research team has released FastFrames-S and FastFrames-L, two video-generation models designed to respond quickly enough for interactive applications. FastFrames-S has 1.3 billion parameters and generates a frame in about 36 milliseconds on one NVIDIA B200 GPU, according to Apple. FastFrames-L has 14 billion parameters and reaches 45 milliseconds on an eight-B200 node. Both target 480p output. (Project page)
The release includes weights, code and evaluation resources. It is a research release, with no announced integration into an Apple consumer product. The accompanying paper, FastFrames: Video Generation at Interactive Latency, was submitted to arXiv on November 17, 2026.
Two Model Sizes, Different Hardware Requirements
Apple describes both models as autoregressive, two-step video diffusion models. Autoregressive generation lets them build video sequentially, incorporating previous frames and new inputs as generation proceeds. That makes the approach useful for camera navigation or other interactions where the next image must reflect a recent action.
The project targets generation latency below 50 milliseconds. Apple’s published results show that the hardware needed to reach that threshold varies considerably:
| Model | Parameters | NVIDIA B200 GPUs | Reported generation latency |
|---|---|---|---|
| FastFrames-S | 1.3 billion | 1 | 36 ms |
| FastFrames-S | 1.3 billion | 8 | 14 ms |
| FastFrames-L | 14 billion | 1 | 158 ms |
| FastFrames-L | 14 billion | 8 | 45 ms |
These are Apple-reported measurements, taken at 480p in the project’s steady-state evaluation setup, not independent measurements. They describe generation latency, so readers should not treat them as guarantees covering application startup, networking, input handling and display delays. (Benchmark conditions and results)
The single-GPU comparison matters most for understanding the hardware tradeoff. FastFrames-L’s 158-millisecond result corresponds to roughly 6.3 generation responses per second, calculated as 1,000 divided by 158. Its headline 45-millisecond result requires eight GPUs.
I’d treat FastFrames-S as the more approachable starting point for a latency-focused experiment. It reaches Apple’s interactive threshold on one accelerator. The larger model’s appeal is visual quality: Apple reports improvements in realism, lighting, composition and text legibility, but reaching comparable responsiveness requires substantially more hardware.
Apple also lists Causal Forcing++ at 272 milliseconds on one B200. Against FastFrames-S at 36 milliseconds on the same GPU count, that is approximately a 7.6-fold latency reduction, consistent with Apple’s nearly eightfold claim.
How FastFrames Reduces Generation Time
Two-step diffusion means the generator uses two denoising evaluations to produce its next output. It does not mean the model generates only two video frames.
FastFrames also changes work surrounding the generator. Apple’s profiling found that conventional video decoding and key-value cache updates could account for more than half of end-to-end inference latency. The decoder turns a compressed internal representation into visible pixels; the cache retains information used while generating subsequent frames.
FastFrames replaces those expensive components with a lightweight pixel decoder and a cache-refinement network, alongside systems optimizations for diffusion sampling. (Research explanation)
The released FastFrames-S profile shows where the time goes afterward:
- First diffusion-transformer evaluation: 16.1 ms
- Second diffusion-transformer evaluation: 16.2 ms
- KV-cache refinement: 2.2 ms
- Pixel decoding: 0.9 ms
Those measurements total 35.4 milliseconds at 480p with a full cache, close to the paper’s rounded 36-millisecond result. (Component timings)
The useful implication is that FastFrames has shifted the bottleneck. Its two transformer evaluations consume 32.3 milliseconds, about 91% of the listed runtime. Based on that published breakdown, making the pixel decoder faster would now deliver only a small overall gain. Further large reductions would need to address the generator itself or how its computation runs.
Existing Models Provide the Foundation
FastFrames builds on released video-generation research instead of training every component from scratch.
For FastFrames-S, Apple keeps the two-step, frame-wise Causal Forcing++ generator frozen and trains the surrounding networks. FastFrames-L starts from Wan2.1-14B, which Apple distills into a two-step causal model before training its pixel decoder and cache-refinement network.
Apple says it releases the large model’s intermediate checkpoints, including the causal teacher, consistency-distilled model, DMD-distilled model and final generator with its supporting networks. Those checkpoints make the training stages available for researchers to inspect. (Training and checkpoint details)
The release also includes an optimized LingBot 1.3B implementation for camera-conditioned control. Apple reports a 16-fold speedup over LingBot’s official implementation. The project demonstrates keyboard-driven navigation, alongside a separate physics-conditioned interaction demo. These demonstrate the intended use case: generated imagery that can respond as an interaction unfolds. (Project demonstrations)

What Developers Can Download and Run
The FastFrames-S and FastFrames-L pages provide access to the release and document generation, profiling, evaluation and training.
The documented environment is Linux x86-64 with an NVIDIA GPU and a CUDA 12.8-compatible driver. This is not an Apple Silicon setup guide.
After obtaining the code, installing its dependencies and downloading the required FastFrames and third-party checkpoints, the documented profiling commands are:
Frequently Asked Questions
3 questions
1Is FastFrames an Apple Consumer Feature?
No consumer integration is announced in the cited materials. FastFrames is a research release with model weights, code and evaluation resources.