CPU3DAI tools for 3D, video and audio

Apple Releases FastFrames-L — a 14B-Parameter text-to-video Model

On November 16, 2026, the model apple/FastFrames-L was published on HuggingFace with the text-to-video task. It is an autoregressive two-step diffusion model with 14 billion parameters that generates 480p video. The largest weights file — fastframes-l/dit.safetensors — takes up 26.6 GB, the entire repository weighs 161.6 GB, and the format is safetensors. The license is apple-amlr. The author also released a smaller version, FastFrames-S at 1.3B, and an optimized LingBot 1.3B, which, as described by the author, runs 16 times faster than the official model and supports camera control. The code has been released as open source.

What it means

Our video generators section does not yet have a tool based on FastFrames. The closest in terms of the text-to-video task is Maestro: it runs on the MiniMax-H3 model, whose largest weights file is 9.7 GB and whose entire repository is 464.1 GB. FastFrames-L has one noticeably larger weights file, but a smaller total size. Maestro requires 6 to 24 GB of VRAM as described by the author; for FastFrames, the developer provides timings only for the S version on a single GPU, without specifying minimum VRAM requirements. Another generator with open source is LTX-Video: its largest weights file is also 26.6 GB, the same as FastFrames-L, but its task is stated as image-to-video rather than text-to-video. As for licensing, FastFrames-L uses apple-amlr, whose terms are not disclosed on the model page; LTX-Video limits commercial use to a revenue threshold, while Maestro's code is permitted for non-commercial use only. For now, FastFrames is not built into any tool in our reference, so it is too early to compare it with ready-made local studios. What is of interest to the reader is the combination of "14B, 480p, interactive latency" and whether integration into existing generators will appear.
FastFrames-L — text-to-video model Autoregressive two-step diffusion · 14B parameters · 480p Input Text prompt Step 1 · Diffusion Frame generation Step 2 · Autoregression Frame linking 480p Video Comparison with peers FastFrames-L 14B parameters · 480p dit.safetensors: 26.6 GB Full repository: 161.6 GB License: apple-amlr text-to-video Maestro MiniMax-H3 Weights file: 9.7 GB Repository: 464.1 GB VRAM: 6—24 GB text-to-video LTX-Video Weights file: 26.6 GB Commercial use limited by revenue threshold Open source image-to-video
How the method works. The diagram was drawn based on this news note.

See also