Mac mini M6 24GB Reality Check: Not Enough RAM to Run Local LLMs

An Impulse Buy Gone Wrong

Many people rushed to buy Apple's newly released M6 Mac mini specifically to run large language models locally. One Bilibili creator was among the first to grab a Mac mini M6 24GB unit from a physical store — thinking he'd dodged the scalper markup — only to get blindsided by a far more insidious problem: not enough memory.

The core miscalculation was in estimating macOS system memory usage. He relied on old assumptions: macOS uses about 8GB at boot, leaving 16GB free from a 24GB pool — plenty to run models at the 3.8B or even 20B parameter scale. That math completely falls apart on the new system.

The New OS Eats RAM, Slashing Usable Space

The crux of the issue is the OS version difference. The creator compared his M1 machine running macOS 26, which used about 9.1GB at boot, against the brand-new Mac mini M6 running macOS 27 — which starts at 12GB and climbs to 14GB under actual use.

24GB minus 14GB leaves roughly 10GB of usable memory. That 10GB isn't enough to load the models he wanted to run. In testing, he tried launching Qwen 3.8B (repeatedly referenced as "3.827B" in the video), typed a simple "Hello," and got an out-of-memory error. The model never got off the ground.

This highlights an easy-to-overlook reality: macOS isn't like Windows, which idles at just 2–3GB. Apple's OS claims a large chunk right out of the gate, and that chunk grows with each new version. Applying old estimates to a new OS version is a reliable way to get burned.

The steady climb in macOS memory usage is closely tied to the layers of functionality Apple has been stacking at the system level in recent years. Spotlight indexing, iCloud sync daemons, Apple Intelligence's local model caches (on supported hardware), Metal Performance Shaders precompilation caches, and the expanded roster of background agents introduced in macOS 26/27 all take up permanent residence in RAM after boot. Apple Intelligence in particular runs local small models for features like writing assistance and summarization — features that simply didn't exist in older OS versions and now occupy several gigabytes on their own. This means that using a single memory estimate across major OS versions introduces compounding error with every release. When evaluating usable memory on any new Mac, the most reliable approach is to reference real-world measurements from the same OS version — not figures from older hardware or older system releases.

What to Buy, and When to Pull the Trigger

The creator's advice is unambiguous: anyone buying a Mac mini to run local LLMs should write off the 24GB model. At minimum, wait for the 32GB version.

Apple's launch strategy also drew criticism. The M6 Mac mini launched with only 24GB configurations available, with 32GB nowhere to be found — seemingly a move to probe both consumer appetite and scalper interest. The silver lining: scalpers didn't manage to inflate prices significantly, and supply has been relatively stable. The bad news: the 32GB model that actually makes sense for running models hasn't been released at all, and prospective buyers are looking at mid-October at the earliest.

His bottom line: if you haven't ordered yet, hold off. If you ordered and it hasn't shipped, consider waiting and redirecting toward the 32GB version. Don't force a 24GB unit to do a job it can't handle.

Even the Flagship Disappoints: The Ceiling on Throwing Hardware at the Problem

The video also references another creator's hands-on with the Mac Studio M6 Ultra with 256GB of unified memory — a machine that runs roughly 60,000 RMB (over $9,000 USD).

The verdict there is equally deflating: after a full day of use, that creator concluded it was "too expensive and not worth it," and fell well short of any vision of "unlimited compute." Generating a 5-second video took about 18 minutes — productivity value far below expectations. It unlocks the ability to run things locally, but it's not a genuine productivity tool.

If even a 256GB, $9,000+ machine can't deliver ideal results, it points to a ceiling on the brute-force-hardware approach to running models on the Mac platform. For ordinary users eyeing a MacBook Air or M4 mini, the idea of running large models locally is best abandoned early.

Video generation tasks place extreme demands on local hardware in ways that are fundamentally different from text inference. Text-based LLMs are primarily bottlenecked by memory bandwidth and capacity — inference mostly involves matrix multiplication in relatively regular computational patterns. Local video generation models (such as those based on architectures like Wan, CogVideoX, or similar), beyond needing to load tens of gigabytes of parameters, must run diffusion denoising iterations (typically 20–50 steps) across every frame, with each step involving extensive temporal convolution and attention computation. The compute demand is one to two orders of magnitude higher than a language model of equivalent parameter count. Generating a 5-second clip in 18 minutes implies a real-time ratio of roughly 1:216 — which isn't unusual for consumer hardware. Even PC workstations with high-end GPUs typically can't generate video locally at anything close to real-time. So "even 256GB isn't enough" is less a verdict on the Mac platform specifically and more a reflection of the universal ceiling for consumer-grade local video generation today.

Mac vs. Gaming PCs: The Flexibility Gap

The creator also made a pointed practical comparison. The Mac mini costs around $1,000+ USD — close to a gaming PC build — and achieves roughly 10 tokens per second inference speed, a number he said he expected and doesn't consider bad.

But the flexibility gap is enormous. Mac's unified memory can't be expanded. If you don't have enough, you don't have enough — no workaround exists. A gaming PC, by contrast, can fall back on system RAM when VRAM runs out. Even if your GPU is a 12GB RTX 5070 or a 16GB 5060 Ti / 5070 Ti, VRAM overflow can spill into system memory — a modest additional cost that solves the hard "can't run at all" problem.

This is the root cause of the Mac mini M6's failure here: a closed unified memory architecture means buying the wrong configuration leaves you with no recourse. In memory-sensitive workloads like local LLM inference, headroom matters more than peak performance.

Apple's Unified Memory Architecture (UMA) integrates the CPU, GPU, and Neural Engine onto a single SoC, with all compute units sharing one memory pool rather than splitting between system RAM and dedicated VRAM as in a PC. The advantage is extremely high bandwidth (M6 chips can reach hundreds of GB/s) and no cross-bus data copying when passing data between processors — theoretically well-suited for large model inference, which constantly shuffles enormous parameter sets around. The tradeoff is that memory capacity is soldered in at the factory and cannot be upgraded like desktop DIMM slots. PC-based LLM inference typically relies on NVIDIA's CUDA ecosystem, which allows model layers to be offloaded to system RAM when VRAM is insufficient — performance drops sharply, but the model runs. Mac's unified memory architecture has no equivalent fallback layer. When available space runs out, the model errors out immediately with no middle ground.

Key Takeaways for Prospective Buyers

If you're considering a Mac mini M6 for local model inference, keep these points in mind:

• macOS 27 consumes about 12GB at boot and climbs higher in practice — 24GB models have critically insufficient usable space
• Wait for the 32GB version at minimum; Apple hasn't released it yet
• If you're on a tight budget or a light user, don't expect to run large models on an Air or M4 mini
• For scenarios where memory flexibility matters, gaming PCs offer better cost-effectiveness and a wider margin for error

The real lesson from this debacle isn't that the M6 chip is inadequate — it's a mismatch between memory configuration and system overhead. When buying any device with unified memory architecture, always calculate usable space against real-world measurements from the actual OS version it ships with, not figures inherited from older hardware and older systems.