FastFrames-S is a 1.3B-parameter autoregressive video diffusion model from Apple that generates 480p video at interactive latency. It runs at 36 ms per frame, about 28 responses per second, on a single NVIDIA B200 GPU. Text-to-video output is 832x480, 81 frames at 16 fps. Weights are on Hugging Face under a non-commercial research license.
A workable 1.3B-parameter dense video generator from Apple. Treat the modality benchmarks above as the leading indicator of fit — composite scoring across modalities is still maturing. Newly released, so production-readiness is still being shaken out.
Generated from this model’s benchmarks and ranking signals. Editor reviews refine it over time.
Access model weights, configuration files, and documentation.
No benchmark data available for this model yet.
The top devices for this model at 4-bit, ranked by fit and speed.
| Device | Grade | VRAM |
|---|---|---|
| ACEMAGIC M1A Pro (i9-13900HK + ARC A770)ACEMAGIC | SS | 1.3 GB |
| Acer Veriton GN100 AI MiniAcer | SS | 1.3 GB |
| AI TOP ATOM ATAGB10-9003Gigabyte | SS | 1.3 GB |
| AMD Instinct MI300XAMD | SS | 1.3 GB |
| AMD Instinct MI325XAMD | SS | 1.3 GB |

Explore the Provider
Aggregate stats, leaderboard, release timeline, and benchmark coverage across every Apple model we track.
Benchmark data sourced from: Hugging Face, Arena.ai, Artificial Analysis.
FastFrames-S 1.3B is an autoregressive video diffusion model developed by Apple designed to solve the critical latency bottleneck in generative video. While standard video diffusion systems take seconds or minutes to produce a short clip, FastFrames-S generates 480p video (832x480 resolution, 81 frames at 16 fps) at interactive speeds. On a single NVIDIA B200 GPU, the model achieves a per-frame latency of 36 ms, delivering approximately 28 responses per second (RPS). This puts autoregressive video generation into sub-50 ms interactive territory for the first time.
Built upon foundations from Wan2.1-T2V-1.3B and Causal-Forcing, FastFrames-S targets workflows where immediate visual response is mandatory, including real-time camera manipulation and interactive physics simulations. Apple released the model weights on Hugging Face under a non-commercial research license, providing researchers and systems engineers with a compact local AI model with 1.3B parameters that challenges the status quo of heavy offline video rendering pipelines.
For local practitioners, FastFrames-S 1.3B occupies a unique space. Rather than requiring multi-GPU server clusters, its compact parameter footprint makes it one of the few video diffusion engines capable of running on individual workstations. It provides an efficient entry point for teams exploring camera-conditioned interaction, real-time world models, and reactive visual agents.
FastFrames-S 1.3B uses a dense transformer architecture containing 1.3 billion parameters. Unlike standard multi-step diffusion pipelines that iteratively denoise full video volumes, FastFrames-S is structured as an autoregressive, two-step diffusion model. It generates visual tokens sequentially across time while conditioning each step on previous frame history.
Apple's profiling of standard autoregressive video pipelines revealed that non-transformer operations account for more than half of total inference latency. In particular, heavy Variational Autoencoder (VAE) decoding and unoptimized Key-Value (KV) cache updates typically stall hardware execution. FastFrames-S replaces these standard bottlenecks with lightweight alternatives:
Combined, these pipeline components bring end-to-end frame synthesis down to roughly 35.4 ms to 36 ms on flagship hardware. The model generates standard outputs at 832x480 resolution with temporal coherence maintained across 81 frames at 16 frames per second.
FastFrames-S 1.3B shifts text-to-video from an asynchronous batch job to an interactive medium. Because frames are produced sequentially with low latency, the model can accept steering signals mid-generation:
To run FastFrames-S 1.3B locally, your setup must prioritize memory bandwidth and low kernel dispatch overhead. Unlike standard text LLMs that run through simplified runtime wrappers like Ollama or llama.cpp, FastFrames-S is an autoregressive diffusion pipeline that executes via PyTorch and CUDA.
Because the model has only 1.3 billion parameters, the raw weights require minimal storage. FastFrames-S 1.3B VRAM requirements vary primarily based on the frame cache and generation context rather than weight storage alone:
A single GPU with 12 GB to 16 GB of VRAM provides sufficient headroom for local execution. For developers considering how to run a 1.3B model on consumer GPUs:
For users seeking the best quantization for FastFrames-S 1.3B, the standard distribution runs in BF16/FP16. Because the model is already compact at 1.3B parameters, weight quantization to INT4 (like standard Q4_K_M formats used in text LLMs) yields diminishing returns for memory savings while potentially degrading high-frequency spatial detail in diffusion sampling. FP8 (specifically FP8 E4M3) is the recommended target if you need to fit the generation pipeline onto sub-8 GB hardware.
The reference implementation requires Linux x86-64, an NVIDIA GPU, and CUDA 12.8. You can initialize the environment using uv:
1# Clone the repository and install dependencies2git clone https://github.com/apple-aiml-research/ml-fastframes3cd ml-fastframes4uv sync56# Run the latency profiling benchmark7uv run python evaluation/latency/benchmark.py ff-s89# Launch local generation10uv run python scripts/generate.py --model ff-s --checkpoint-dir checkpoints --prompt "A drone flying smoothly through a canyon"
In terms of FastFrames-S 1.3B performance, throughput is measured in latency per frame or responses per second rather than standard FastFrames-S 1.3B tokens per second. While enterprise B200 hardware reaches 28 RPS (36 ms per frame), consumer hardware like an RTX 4090 achieves around 55 ms to 80 ms per frame, which remains responsive enough for fluid prototyping.
Evaluating FastFrames-S requires looking at both its base architecture and competing video diffusion models in the 1B to 2B parameter class.
Wan2.1-T2V-1.3B serves as one of the architectural foundations for FastFrames-S. However, standard Wan2.1 relies on conventional multi-step diffusion sampling (often requiring 30 to 50 steps per sequence) and a full-scale VAE decoder.
Apple released FastFrames-S alongside its larger sibling, FastFrames-L (14B parameters).
Cheapest current cloud rentals with at least 1 GB VRAM, refreshed hourly.
Advertising disclosure: we earn commissions when you shop through the links below.
| Option | Cost / GPU-hour |
|---|---|
NVIDIA GeForce RTX 3060Vast.ai · Spot · 12 GB VRAM | $0.06 |
NVIDIA GeForce RTX 3060Vast.ai · On-Demand · 12 GB VRAM | $0.06 |
NVIDIA GeForce RTX 5070Vast.ai · Spot · 12 GB VRAM | $0.07 |
NVIDIA GeForce RTX 5060 TiVast.ai · Spot · 16 GB VRAM | $0.08 |
NVIDIA RTX A4000Vast.ai · Spot · 16 GB VRAM | $0.08 |
Per-GPU rate across RunPod, the Vast.ai marketplace, DigitalOcean, and Vultr.
Spot tier is interruptible. Plan for restarts when comparing against on-demand prices.