FastFrames-S on Mac mini: 30.4s for 5s Video, 23GB Peak Memory – Is 64GB Enough?
The article benchmarks Apple's FastFrames-S 1.3B video model ported to MLX on a 64GB M5 Pro Mac mini, revealing 30.4 seconds to generate 5 seconds of video at 376ms per frame with 23GB peak memory, comparing it to the official 36ms/frame on NVIDIA B200, and analyzes memory requirements for different Mac configurations while noting the non-commercial license.
Introduction
Apple open-sourced the FastFrames video model on November 16, releasing two sizes: S (1.3B parameters) and L (14B). The official implementation targets Linux with NVIDIA GPUs and CUDA 12.8, reporting a single-frame latency of 36 ms on a single B200 GPU. A community port to Apple's MLX framework now allows running FastFrames-S on Mac hardware. Pratik Gandhi tested the MLX port on a 64GB M5 Pro Mac mini and shared concrete performance numbers.
Hard Numbers from the Tweet
The following measurements come directly from the author's self-reported tweet; they are not official Apple benchmarks nor third-party standardized tests.
Model: FastFrames-S, 1.3B parameters.
Output: 832×480 resolution, 81 frames at 16 fps, approximately 5 seconds of video – matching Apple's official sample settings.
Generation time: 30.4 seconds total, yielding 376 ms per frame.
Peak memory: ~23 GB on a 64GB M5 Pro Mac mini.
Relative to real-time: 30.4 s ÷ 5 s ≈ 6.08× slower than real-time playback.
The key takeaway is not merely that a Mac can run a video model – that was expected – but that the tweet provides two decision-critical figures: peak memory consumption and per-frame latency.
Official 36 ms vs. Mac 376 ms: Understanding the Gap
Two distinct axes must be separated:
Total generation time: The video plays in 5 s; the Mac mini needs 30.4 s to produce it – roughly 6× slower than real-time, as the author notes.
Single-frame latency: Apple's official B200 run achieves 36 ms/frame; the MLX port on Mac mini measures 376 ms/frame. The ratio 376 ÷ 36 ≈ 10.4× (estimated).
This comparison crosses different hardware (data-center B200 vs. desktop Mac mini) and different software stacks (highly optimized CUDA code vs. community MLX port). The 10.4× figure is only an order-of-magnitude reference, not a general “Mac is 10× slower than B200” conclusion. Consequently, FastFrames' “interactive latency” claim currently holds only in the specific NVIDIA B200 environment; on Mac the model can generate video but not at interactive speeds.
23 GB Peak Memory Dictates Your Memory Tier Choice
Based on the observed 23 GB peak, the author recommends:
16GB / 32GB: Do not attempt. 16GB is far below the peak; 32GB leaves almost no headroom for the OS and other processes. This is a recommendation + extrapolation, not a direct test on those configurations.
64GB: The tested tier – runs comfortably with margin. Based on Apple's official pricing, a base 24GB/512GB Mac mini starts at ¥12,999; upgrading to 64GB adds ~¥7,500, totaling ~¥20,499.
128GB+: More comfortable headroom, but will not make generation faster . On the same chip, memory capacity determines whether the model fits; speed is governed by compute and bandwidth – true for video models as well.
License reminder: The model weights are released under the Apple Research Model License, which permits research use only – no commercial usage. Verify the license before building any commercial product around it.
How to Judge Whether Your Machine Can Run It
The site referenced at the top of the article (workbuddy.link) currently indexes 201 models, 126 hardware configurations, and 12,609 speed data points – but its focus is language models ; FastFrames and other video models are not included. The evaluation method, however, is identical:
Fit check: Can the model's peak memory (23 GB for FastFrames-S) fit in your unified memory?
Speed check: What is the actual per-frame latency on your hardware (376 ms/frame on M5 Pro) versus the marketing number (36 ms/frame on B200)?
Separating these two steps prevents being misled by “36 ms real-time generation” headlines that assume a very different hardware stack.
Final Question
If generating a 5-second video locally takes 30 seconds, is that acceptable for your workflow? Or would you prefer to wait longer for the higher-quality 14B version? Share your machine's memory and your tolerance for wait time in the comments – the author will help estimate whether upgrading to 64GB is worthwhile.
Signed-in readers can open the original source through BestHub's protected redirect.
This article has been distilled and summarized from source material, then republished for learning and reference. If you believe it infringes your rights, please contactand we will review it promptly.
How this landed with the community
Was this worth your time?
0 Comments
Thoughtful readers leave field notes, pushback, and hard-won operational detail here.
