How to unlock more VRAM on your Mac for local AI models

If you’ve tried running large language models locally on a base Apple silicon Mac, you’ve probably hit a wall fast. A model that looks like it should fit into 16GB of RAM simply refuses to load, or your system grinds to a halt. The good news: your Mac has a hidden memory limit you can tweak to squeeze more VRAM out of the same hardware.

Why your 16GB Mac can’t load a 11GB model (by default)

Apple silicon Macs use unified memory. That means the CPU and GPU share the same pool of RAM instead of having separate system RAM and VRAM. On paper, a 16GB machine plus an 11GB model sounds fine. In practice, it often fails.

Tools like LM Studio make this obvious. On a 16GB Mac, LM Studio might show something like:

• RAM: 16GB
• VRAM: ~11.8GB usable for models

That VRAM number is the real limiter. When you try to load a model such as GPT-OSS-20B (around 11.28GB on disk), LM Studio also needs extra memory for context, overhead, and runtime operations. The estimated total might jump to 12GB+ — more than the default VRAM budget macOS is willing to give the GPU.

Even though you technically have 16GB, macOS must reserve some of it for:

• The operating system kernel
• WindowServer and UI rendering
• GPU drivers and I/O buffers
• Compressed memory and background tasks

You can see many of these in Activity Monitor under the Memory tab. macOS is very good at juggling memory, but it still enforces limits so the GPU can’t starve the rest of the system.

The hidden GPU memory limit macOS uses

Behind the scenes, macOS has a setting that controls how much memory the GPU is allowed to lock for itself. Think of it as a cap on how much of your unified memory can be treated as VRAM.

By default, this limit is set to 0, which doesn’t mean “no limit” — it means “let macOS decide”. In practice, that usually works out to roughly 75% of your total memory. On a 16GB Mac, that’s around 12GB; on a 512GB Mac Studio, it’s about 384GB.

The trick: you can override this default and tell macOS to allow the GPU to lock more (or less) memory, effectively changing how much VRAM tools like LM Studio see.

Checking your current GPU memory lock limit

You’ll need to use the Terminal and an admin account. This involves sudo, so only proceed if you’re comfortable with command-line changes and understand there is some risk.

First, run the command that reads the current GPU memory lock limit. It will prompt for your password and usually returns 0 on a fresh system, meaning the default behavior is active.

When it’s at 0, macOS calculates the limit automatically (around 75% of total RAM). That’s why LM Studio initially shows something like 11.8GB VRAM on a 16GB machine instead of the full 16GB.

Changing the limit: powers of two and safe values

You can set this limit manually in megabytes. Memory-related values in computing are usually powers of two (2, 4, 8, 16, 32, 64, and so on), and this setting follows the same pattern.

Examples of values you might use on a 16GB Mac:

• 8192 MB (8GB)
• 14336 MB (~14GB)

After setting a new value and restarting LM Studio, you’ll see the VRAM number change in its hardware settings. For instance, setting 8192 MB may make LM Studio report around 8GB of VRAM instead of ~11.8GB, proving that the setting is actually controlling the GPU’s memory budget.

Pushing it too far (and why you shouldn’t)

You can also push this limit very high — for example, telling macOS to allow roughly 16GB of VRAM on a 16GB machine. LM Studio may then show something like 15.6GB VRAM, and surprisingly, a big 20B-parameter model might actually load and run.

But there’s a catch: your memory pressure in Activity Monitor will spike into orange or even red. When that happens, macOS is under heavy stress, aggressively compressing memory and juggling processes to keep the system alive. It might still work, but it’s unstable and not recommended for everyday use.

In short: just because you can set the limit to nearly all your RAM doesn’t mean you should. Leave some headroom for the OS and background tasks.

A more realistic setting for 16GB Macs

For a 16GB Apple silicon Mac, a more reasonable value is around 14GB of VRAM. In megabytes, that’s 14336 MB. This still isn’t perfectly “safe”, but it’s a practical compromise if your main goal is to run local models and you’re not doing much else on the machine at the same time.

With the limit set to 14336 MB and LM Studio restarted, you’ll typically see:

• VRAM: ~14GB available to models
• Enough leftover memory for macOS and a few background processes

At that point, a model like GPT-OSS-20B can load successfully, and you can generate text at a decent speed (for example, around 16–17 tokens per second) without immediately pushing memory pressure into the danger zone.

How this scales on bigger Apple silicon machines

This same trick scales up nicely on higher-end Apple silicon desktops like Mac Studio or Mac mini with large memory configurations. For example, if you have a 512GB Mac Studio, the default 0 setting will cap GPU-locked memory at roughly 75% of that, which may still be too conservative if you’re running multiple large models or a cluster of machines.

By manually increasing the GPU memory lock limit, you can dedicate more of that huge unified memory pool to VRAM and run very large models or multiple instances in parallel. This is especially useful if the machine is running headless (no display, no heavy UI), and its main job is serving local AI workloads.

If you’re interested in squeezing big models into limited VRAM, you may also find it useful to learn how to optimize model formats. For example, you can see how a 3D AI model is tuned to run locally in this guide to running Trellis 2 3D AI on just 6GB VRAM.

Watching memory pressure while you experiment

Whenever you change this GPU memory limit, keep Activity Monitor open on the Memory tab. The key things to watch are:

• Memory used: how much of your total RAM is currently in use
• Memory pressure: the colored graph at the bottom

Green means you’re fine, orange means you’re pushing it, and red means your system is struggling. If you see sustained orange or red while running a model, consider lowering the GPU memory limit or using a smaller model.

When this tweak makes sense (and when it doesn’t)

This hidden setting is most useful when:

• You’re running local LLMs on a small-memory Mac (8–16GB)
• You’re using tools like LM Studio that clearly show VRAM usage
• The machine is mostly dedicated to AI workloads (especially headless servers)

It’s less useful or riskier when:

• You multitask heavily (browsers, video editing, games, etc.)
• You rely on rock-solid stability more than raw model size
• You’re not comfortable reverting system-level tweaks if something breaks

If you’d rather offload more of the heavy lifting to smarter orchestration instead of pushing your hardware to the edge, you might be interested in building higher-level systems around your models. For example, you can learn how to coordinate tools and models into a personal AI stack in this guide on building your own agentic operating system.

Final thoughts

Apple’s unified memory architecture is powerful, but the default settings are conservative to keep your system stable. By carefully increasing the GPU memory lock limit, you can unlock more VRAM on your Mac and run larger local language models than you’d expect from the spec sheet.

Just remember: treat this like overclocking. Start with modest values, watch memory pressure closely, and always leave some room for macOS to breathe. Done right, this small tweak can turn even a modest MacBook Air or Mac mini into a surprisingly capable local AI machine.