Apple Releases FastFrames-S — a text-to-video Model with Interactive Latency
On November 16, 2026, the model apple/FastFrames-S for text-to-video generation appeared on HuggingFace. It is the smaller version in the FastFrames family: it contains 1.3 billion parameters, works as a two-step autoregressive diffusion model, and generates video at 480p resolution. The weights are published in safetensors format, the largest file is kv_refiner.safetensors at 0.6 GB, and the entire repository takes up 1.6 GB. The license is apple-amlr. The code is released as open source.
What it means
FastFrames-S is a standalone model, not an update to the tools in our reference. There is no direct overlap with Maestro, Imagination Studio or LTX-Video yet: none of them uses FastFrames as a model.
As for hardware requirements, Apple lists only general conditions: Linux x86-64, an NVIDIA GPU and a driver with CUDA 12.8 support. The developer does not specify the exact amount of VRAM. For comparison: LTX-Video claims it runs from 1 GB of VRAM, Maestro from 6 GB, Imagination Studio from 8.3 GB. FastFrames-S, with its 1.6 GB of weights, looks noticeably lighter in size, but without data on peak memory consumption a direct comparison is not possible.
The model's main feature is its focus on interactive latency: the model card lists component timings, where a single DiT call takes about 16 ms on one GPU at 480p resolution. This positions FastFrames-S as a tool for fast generation rather than for maximum frame quality.
For now the model has only just been released, and it does not appear in our tool fact sheets. If any developers of local studios add FastFrames-S support, that will be a separate update to the tool page.How the method works. The diagram is drawn based on this news note.