Blog · AI Agent News · AI video VIEN

Apple Releases FastFrames: 35.4 ms AI Video on One GPU

Apple open-sourced FastFrames (14B & 1.3B), cutting 480p AI video latency to 35.4 ms on one GPU. Here are the benchmarks and hardware limits.

Apple Releases FastFrames: 35.4 ms AI Video on One GPU
Images: 极客湾Geekerwan / Wikimedia Commons (CC BY 3.0)
In this article
  1. How do FastFrames-S (1.3B) and FastFrames-L (14B) compare?
  2. How Apple broke the 35.4 ms bottleneck
  3. What this means for local video creators
  4. My take

Generating a short AI video clip usually means waiting long enough to grab a coffee before seeing the first frame, yet in November 2026 Apple AIML researchers released FastFrames, an open-weight model family that generates 480p video at 35.4 ms step latency on a single GPU. Rather than shrinking the diffusion backbone alone, the team targeted two often-overlooked stages: pixel decoding and KV cache updates.

Bar chart comparing single-step 480p video generation latency across models
FastFrames-S reaches 35.4 ms on a single NVIDIA B200 GPU, running 8x to 20x faster than prior autoregressive video models

How do FastFrames-S (1.3B) and FastFrames-L (14B) compare?

Apple released two two-step autoregressive diffusion variants aimed at distinct workloads, both generating 832 × 480 clips (81 frames at 16 fps). Alongside them, the team published an optimized build of LingBot-World v2 1.3B that runs 16x faster than the upstream implementation for keyboard-controlled camera navigation.

  • FastFrames-S (1.3B parameters): Freezes a two-step frame-wise Causal Forcing++ generator and trains only the surrounding pixel decoder and KV refinement module. On one NVIDIA B200 GPU, it hits 35.4 ms per four-frame RGB step (dropping to 14 ms across eight GPUs), opening the door to real-time interactive worlds.
  • FastFrames-L (14B parameters): Distilled from Wan2.1-T2V-14B, trading raw speed for noticeably sharper lighting, shadows, composition, and in-scene text rendering. It runs at 158 ms per step on one B200 GPU (45 ms across eight GPUs), matching Krea-Realtime quality while running over 10x faster.
ModelVBench TotalVBench Quality1-GPU Latency (B200)8-GPU Latency (B200)
CausVid83.1884.15——
Self-Forcing84.1084.84717 ms—
Causal Forcing++84.1185.09272 ms—
FastFrames-S (1.3B)84.2385.3035.4 ms (paper: 36 ms)14 ms
FastFrames-L (14B)84.4485.48158 ms45 ms

How Apple broke the 35.4 ms bottleneck

Most speed-focused video papers try to cut diffusion denoising steps down to one or two passes, yet Apple's component profiling showed that two surrounding operations—decoding latents to pixels and updating the KV cache—were eating a huge share of wall-clock latency.

flowchart LR
  A[2-step DiT: 32.3 ms]:::data --> B[KV refinement: 2.2 ms]
  B --> C[Pixel decoder: 0.9 ms]:::key
  C --> D[4 RGB frames at 480p]
  classDef key fill:#1d1a10,stroke:#ffd23f,color:#ffd23f,stroke-width:1.6px
  classDef human fill:#181817,stroke:#e8e6df,color:#e8e6df,stroke-dasharray:5 3
  classDef data fill:#1c1c1b,stroke:#8d8b85,color:#cfcdc6
  classDef muted fill:#181817,stroke:#3d3d3a,color:#8d8b85
Diagram: Four stages in a single FastFrames-S generation step

First, VAE decoding—the step that converts compressed latent representations back into visible RGB pixels—was replaced with a compact streaming pixel decoder. In FastFrames-S, decoding a latent frame takes just 0.9 ms, eliminating the stall caused by calling a heavy full-frame VAE at every interactive chunk.

Metric card breaking down the four component timings inside 35.4 ms
On one NVIDIA B200 GPU, FastFrames-S spends 32.3 ms on two DiT passes, 2.2 ms on KV refinement, and 0.9 ms on pixel decoding

Second, KV cache updates—which keep characters and camera geometry consistent across generated frames—no longer require a dedicated clean DiT pass as in Causal Forcing++. Instead, a compact KV refinement network updates cache states in 2.2 ms. Combined with the two DiT evaluations (16.1 ms at t=1000 and 16.2 ms at t=833), the entire pipeline lands at 35.4 ms.

Also readTurning a 2D Painting Into a 3D Scene With Opus 5.5 in toEdit

What this means for local video creators

Interactive latency sounds immediately usable on a desk rig, yet creators running 8–16 GB consumer GPUs or Apple Silicon Macs cannot deploy these checkpoints out of the box just yet. Apple's numbers come from data-center hardware under a narrow software stack.

Before local video editors see millisecond feedback on everyday machines, the open-source ecosystem will need quantized checkpoints (such as FP8 or INT4) or Apple Silicon MLX ports. Even so, the release proves that stripping away VAE and KV bottlenecks matters just as much as shrinking model weights—especially as rising memory and hardware costs for local AI push creators to squeeze more out of every gigabyte, while NVIDIA and CUDA remain central even for research shipped by Apple.

My take

For editors and solo creators, the worst friction in AI video today is not rendering the final polished clip—it is waiting one or two minutes per prompt just to test a camera angle or pacing idea, burning credits every time the framing misses.

Once frame latency drops into the tens of milliseconds, video generation shifts from "submit and wait" to live interactive storyboarding. Consumer cards and Apple Silicon Macs still need quantization and runtime ports to catch up, but Apple's profiling roadmap shows clearly how local AI video stops feeling like a slow render queue.

Sources

  1. FastFrames: Video Generation at Interactive Latency (Project Page) Apple AIML Research
  2. apple-aiml-research/ml-fastframes GitHub
  3. apple/FastFrames-S & apple/FastFrames-L Hugging Face
  4. Gene Chou (Apple AIML) — Publications Gene Chou
  5. X @_akhaliq — Apple FastFrames-L 14B & FastFrames-S 1.3B X

AI videoOpen models

Was this useful?

More on this topic