Video Apple Silicon Benchmarks
How long LTX 2.5 takes to make a clip on Apple Silicon, through the same generate_video call an AI assistant makes: 704 x 1280, five seconds, measured across twenty clips.
Video generation time per clip
Wall-clock seconds, timed through the same tool an AI assistant calls.
- One clip
- One generate_video call at 704 x 1280, 121 frames, which is 5.04 seconds at 24 fps. Timed through the tool, so the figure includes the queue, the final encode and saving the clip to the library.
Apple Silicon Relative Performance Metrics
The same measurements as charts. Each chart has its own scale, and the default model is marked.
At 121 frames the weights are freed after every render and loaded again for the next one, so every clip pays the same load time. The twenty clips measured all fall within 122 seconds of each other, with no fast first run or slow cold start.
Before a render, the app estimated 489 seconds for this size. The measured median was 454.75. On this size the estimate runs long, so a render usually finishes before it says.
A job's elapsed time starts when it is accepted, not when it starts rendering, so a job waiting behind another reports roughly double. Timed from one finished clip to the next, the last clip in the batch took about as long as the first.
Makers
The companies behind these models.
LTX 2.5 model files and download size
Every model GroundSlate offers for download, including ones not measured yet.
How video generation works on a Mac
LTX 2.5 installs as two repos into one folder. The base pack holds the picture decoder, the Gemma-4 text encoder that reads your prompt, the upscalers and the vocoder. The distilled transformer, which makes the frames, installs on top of it. Neither works alone, so they load together, and a generated clip records the pack as its source.
At the size timed here, the weights do not stay in memory between clips. At 121 frames they are freed after every render and loaded again for the next one. So there is no separate warm or cold figure: every clip pays the same load time. Twenty clips timed over five hours varied by only about two minutes.
How these numbers were measured
- Machine
- MacBook Pro, Apple M3 Max, 64 GB unified memory
- Engine
- mlx
- Default model
- mlx-community/ltx-2.5-mlx-q8
- Models locked
- Yes
- Path
- Timed through the generate_video tool, so the figure includes the queue, the final encode and saving the clip to the library.
- Shape
- 704 x 1280, 121 frames, 24 fps. All twenty clips were checked afterwards for size and frame count, and all matched.
- What was timed
- The time between one clip finishing and the next. The queue accepts three jobs at once but renders one at a time, so a job's own elapsed time includes its wait and can roughly double under load.
- Cold against warm
- The same at this size. At 121 frames the weights are freed after every render and loaded again for the next clip.
Questions
- How fast is LTX 2.5 on a Mac?
- On a MacBook Pro, Apple M3 Max, 64 GB unified memory, one clip takes 7m 35s, which is 1m 30s per second of video. Timed through the tool an agent calls.
- Is there a warm run to plan around, or a cold start to avoid?
- At 121 frames the weights are freed after every render and loaded again for the next one, so every clip pays the same load time. The twenty clips measured all fall within 122 seconds of each other, with no fast first run or slow cold start.
- Can the render time the app predicts be trusted?
- Before a render, the app estimated 489 seconds for this size. The measured median was 454.75. On this size the estimate runs long, so a render usually finishes before it says.
- Do renders get slower as a long batch goes on?
- A job's elapsed time starts when it is accepted, not when it starts rendering, so a job waiting behind another reports roughly double. Timed from one finished clip to the next, the last clip in the batch took about as long as the first.
- Which machines are in this dataset?
- MacBook Pro, Apple M3 Max, 64 GB unified memory (Reference Mac).
- Which video model does GroundSlate use by default?
- mlx-community/ltx-2.5-mlx-q8. LTX 2.5, 44.5 GB on disk.