Tasks and inputs
FastFrames-L is listed for Text-to-video, Text → Video. It accepts text and produces video. Use the publisher’s input format and preprocessing for your chosen deployment.
FastFrames-L is Apple's open 14B autoregressive two-step video diffusion model that generates 480p video at interactive latency, distilled from Wan2.1-14B.
The specification score summarizes recorded features and limits. It is not a benchmark result.
Pricing conditions: Open weights; no per-token price is published. Your cost is the GPU time you run it on.
Recorded modality support shows what this model can accept and produce. An undocumented modality is shown as unknown rather than assumed to be unsupported.
Accepts
Produces
Most video generators make you wait for the whole clip. FastFrames generates video frame by frame, fast enough to steer while it plays. Apple's research team released two sizes on 16 November 2026: FastFrames-S at 1.3B parameters and FastFrames-L at 14B, both autoregressive, two-step video diffusion models producing 480p.
The L model trades some of the small one's speed for the quality of a 14B base.
Apple first distilled Wan2.1-14B into a two-step causal model, following the Causal Forcing++ recipe. Then it trained a pixel decoder and a KV-refinement network around it, as it did for FastFrames-S. Every stage is released: the causal teacher, the consistency-distilled model, the DMD-distilled model, and the final generator with its decoder and refinement network. That makes it unusually easy to study.
Apple's profile of FastFrames-S on one GPU at 480p with a full KV cache:
No equivalent breakdown is given for the L model. Measure it yourself.
Text-to-video runs write 832 × 480, 81 frames at 16 fps; pass --model ff-l for the 14B. VBench evaluation scripts reproduce the paper's quality and semantic scores, though the paper itself was still marked as coming soon.
Requirements: Linux x86-64, an NVIDIA GPU and a CUDA 12.8-compatible driver.
The release also includes an optimised LingBot 1.3B, 16x faster than the official one, for camera-controlled interactive scenes in a browser demo.
Weights use Apple's AMLR licence, a research licence. FastFrames code is under the Apple Sample Code License; LingBot-World v2 parts keep their upstream CC BY-NC-SA 4.0 terms.
You are prototyping interactive or game-like video and need to know where latency goes. Generate twenty prompts with S and L, time per-step latency on your GPU, and compare VBench-style quality. If L cannot hold the frame rate you need, S is the fallback. Compare offline quality with the open text-to-video models listed, or browse Apple's other open models on its provider page.
Compare the model on the work you need to complete.
FastFrames-L is listed for Text-to-video, Text → Video. It accepts text and produces video. Use the publisher’s input format and preprocessing for your chosen deployment.
Choose duration, resolution and frame rate together. Compare the cost of an accepted clip, including repeat generations.
The weights are downloadable under Apple AMLR (weights); Apple Sample Code License (code). Check runtime support and size the complete model and working context for your expected concurrency.
These records share a directory category. Check the source card, output type and deployment requirements before treating a model as a replacement.
Genie 4 is Google DeepMind's interactive world model with real-time native audio, available in Project Genie to eligible Google AI Ultra subscribers.
Gemini Omni 1.2 Flash is Google's conversational video generation and editing model, making 3-15 second clips with native audio from 360p to 4K.
pika
Pika H3 Flex is Pika's open LoRA post-training of MiniMax H3: six-step sampling, up to 15 s clips with 32 kHz stereo audio at 480p or 768p, 24 fps.
Tencent
Hy-Nomas is Tencent Hunyuan's open interactive world model: it generates 720p video segment by segment, steered by text and camera controls.
heygen
HeyGen Video 1 generates 5 to 30 second MP4 clips with audio from text, images or reference media, up to 2K at 24 fps. Built on MiniMax H3.
Meta
Muse Video is Meta’s previewed video-generation model with native audio; a public Meta Model API route and price have not been established.