EN AR

Apple releases 14-billion-parameter FastFrames video model, research use only

Tamara khadrah

Key Points

  1. Apple released FastFrames-L, a 14-billion-parameter text-to-video model, plus a 1.3-billion-parameter FastFrames-S version.
  2. Both are autoregressive two-step diffusion models generating 480p video at interactive latency, per Apple's model card.
  3. The license bars commercial products and services, limiting the release to research work.

The latest:

A 14-billion-parameter text-to-video model, FastFrames-L, is now available from Apple alongside a smaller 1.3-billion-parameter sibling, FastFrames-S. Both are autoregressive, two-step diffusion models that produce 480p video at interactive latency, according to the company’s Hugging Face model card. The accompanying license restricts use to research and excludes commercial products and services.

Details:

  • What shipped: The release bundles model checkpoints, generation and evaluation scripts, and an interactive demonstration, according to the model card. A separate script is provided to reproduce the more than 20 videos shown on the project page. Apple did not announce any commercial availability timeline.
  • How FastFrames-L was built: Apple says it distilled Wan2.1-14B into a two-step causal model using the Causal Forcing++ training recipe, then trained a pixel decoder and a key-value refinement network on top of it. The released checkpoints span a causal teacher, a consistency-distilled model, a DMD-distilled model and the final generator.
  • The smaller model: For FastFrames-S, the company kept the two-step, frame-wise Causal Forcing++ generator frozen and trained the surrounding networks instead. The model card names Wan2.1, Self-Forcing and Causal Forcing++ as the earlier projects the FastFrames code and models build on.
  • Output specs: An example generation command for FastFrames-S produces video at 832 by 480 pixels, with 81 frames at 16 frames per second, according to the model card. The same script can be configured to run FastFrames-L instead.
  • The latency numbers: Apple provides a timing breakdown for FastFrames-S at 480p on a single GPU with a full key-value cache totaling 35.4 milliseconds. The two diffusion transformer evaluations account for 16.1 and 16.2 milliseconds, key-value refinement 2.2 milliseconds and the pixel decoder 0.9 milliseconds.
  • Scope of the figures: Those timings cover the 1.3-billion-parameter model only. Apple did not publish an equivalent component-level breakdown for the far larger FastFrames-L, leaving its per-step cost on the same hardware unstated in the release.
  • Hardware requirements: Running the release requires Linux x86-64, an NVIDIA GPU and a CUDA 12.8-compatible driver, plus the FastFrames package and third-party weights. Apple also supplies commands for latency profiling and instructions for reproducing quality, semantic and total scores using the VBench benchmark.
  • Camera control: The package includes an optimized LingBot 1.3B model for camera-conditioned control. Apple says that version is 16 times faster than the official model and allows fluid camera-conditioned control, with the interactive demo startable on either one or eight GPUs.
  • Licensing split: LingBot-World v2 code and weights retain their upstream CC BY-NC-SA 4.0 license, while the remaining FastFrames code falls under the Apple Sample Code License. The combination keeps the full stack outside commercial deployment.

Between the lines:

The release is engineered around speed rather than resolution: a 35.4-millisecond component total at 480p, two-step diffusion and a 16-times-faster camera-control model all point at real-time generation as the target. The research-only license and the absence of published large-model timings keep the headline 14-billion-parameter system a reference point rather than a deployable product.

What’s next

Watch for VBench scores reproduced by outside researchers using Apple’s published instructions, a component-level latency breakdown for FastFrames-L, and any change to the research-only license terms.

What to read next