Macs handle large models well, but video exposes their limits: measured 30x slower. Recently there's been a heated debate in the community—some are spreading the news of "MiniMax H3 running on Apple" as great news, saying it's a major innovation for the Apple ecosystem and that even small machines can now produce AI video. On the other side, a video blogger went straight on the attack, saying it's a scam—Apple is inherently crippled when it comes to running AI video, and no amount of porting will help.

💡 What You Will Learn

Macs handle large models well, but video exposes their limits: measured 30x slower. Recently there's been a heated debate in the community—some are spreading the news of "MiniMax H3 running on Apple"

📜 Table of Contents

• The Bottom Line: It Runs, But So Slowly You Won't Want to Use It
• Real-World Gap: 1.5 Hours vs 2-3 Minutes
• Why Large Models Work but Video Doesn't
• So Is a Mac Worth Buying?
• At the End of the Day, It Depends on Whether Your Machine Is for Working or for Show

Running Large Models on a Mac Is Impressive — Until You Try Video: Measured 30x Slower

There's been a heated debate in the community lately — some people are spreading "MiniMax H3 runs on Apple" as great news, calling it a major breakthrough for the Apple ecosystem and claiming even a mini desktop can produce AI video. On the other side, a video blogger went straight on the attack, calling it a scam and saying Apple is inherently crippled for AI video — no amount of porting will help.

This topic is deeply relevant to me. I use H3 to produce videos every day — I've run more machines, hit more pitfalls, and shipped more finished videos than either side of this argument. So today I'm not picking a side. I'll just lay out the data and conclusions and talk about whether a Mac can actually serve as an AI video production tool. The answer might sting more than either side expects.

The Bottom Line: It Runs, But So Slowly You Won't Want to Use It

Can Apple run H3? Yes. No need to argue about that — the community has already proven it works. The father of Redis (yes, the guy who wrote Redis) built a native Metal inference engine called h3.c, heavily optimized for Apple Silicon. On an M5 Max at 512×512 resolution with 4-step denoising, it can get down to about 3.5 seconds per step, roughly 19 seconds total for a small sample. Someone even got the model running on a 16GB Mac using weight streaming. So "it can't run" doesn't hold up.

But "it can run" and "it's usable for real work" are two different things.

Real-World Gap: 1.5 Hours vs 2-3 Minutes

Let's look at some real numbers from the community — that's the most intuitive way. Someone on Hacker News used an M4 Max 128GB Mac Studio to run H3 in ComfyUI and generate a 15-second, 480p video — just basic text-to-video — and it took an hour and a half.

The same task on an RTX Pro 6000 GPU: two to three minutes.

So how much slower is that? About 30x. I thought I'd misread it at first and checked several times — it's really that order of magnitude.

My own machine is on the NVIDIA side, using a 3-step LoRA approach, and a 124-frame clip takes about 6 minutes. I haven't done a same-config comparison with the Apple setup, but just putting that time gap on the table, I have a clear sense of what it means to talk about "Apple mini desktops taking over AI video production" — this isn't a question of fast vs. slow, it's a question of whether it can be a productivity tool at all.

Why Large Models Work but Video Doesn't

There's a counterintuitive point here worth spelling out: running AI on a Mac is impressive for large language models, but the moment you try video, it falls apart. Apple is catching up — after the M5 generation, running MoE models and Qwen 27B-scale models, the capability has genuinely improved; new architectures like Strata and SGLang are also being ported to Apple. Unified memory's high bandwidth has narrowed the gap between Mac and NVIDIA considerably for text generation tasks.

But video generation follows a different logic. Models like H3 generate video through an iterative denoising process — step by step, computing the image out of noise — and finally passing through a decoder called VAE to turn the compressed latent space into the final image. The core of each step is raw BF16 compute and massive matrix operations, which memory bandwidth alone can't compensate for. Apple chips' actual compute in this area is orders of magnitude behind NVIDIA's flagship GPUs.

Some people attribute this to "Apple's FP32 being cut off at the knees," saying that's why it can't run video. That's half right: Apple's floating-point compute has indeed been heavily reduced, but the real bottleneck in community testing is more about dense attention computation and BF16 linear operation throughput, not any single instruction set. The attribution could be more precise, but the direction is correct — video generation, this "high-intensity per-frame compute" work, is Mac's nightmare mode.

So Is a Mac Worth Buying?

Having written all this, I should be fair: Macs aren't without value. Low power consumption, quiet, stable system — for video editing, coding, and daily office work, a Mac mini starting around 4,000 yuan is still a good machine. Quite a few friends around me bought one for exactly that. Its AI weakness only affects one scenario: "using it for AI production."

So don't get swept up by either extreme. On one side, "major breakthrough, Apple is about to defy the heavens"; on the other, "pure scam, don't buy it." Neither is accurate. The truth is: a Mac can run H3, but as an AI video production tool, it currently doesn't justify the price; its value is in office work and light creative tasks, not in running models to produce videos.

At the End of the Day, It Depends on Whether Your Machine Is for Working or for Show

I make my living from what my machines produce. The only value a device has for me is turning footage into finished videos. The evaluation criterion for "can a Mac run AI video" has been one sentence from start to finish: how fast can it get the video out for you. 1.5 hours for a 480p clip — for me, that's not a question of being crippled or not, it's that the work simply can't get done.

So what about you? If you have a high-spec Mac and have actually installed H3 with your own money, does the speed you're getting match what I've described? Let's talk in the comments.