← BASELOAD English

Local LLMs on a MacBook — What Runs on 8, 16, 32, 64GB Unified Memory

July 8, 2026 · 34 min read 한국어 원문

“Can a MacBook actually run a local LLM?” The answer is decided almost entirely by not which MacBook you have, but how many GB of unified memory it has. Bottom line up front: 16GB is comfortable up to ~14B, 32GB handles ~32B, and 64GB reaches for 70B. What’s interesting is that the MacBook turns out to be a surprisingly strong contender for local AI — because where a gaming GPU hits a wall at 24–32GB of VRAM, a Mac’s unified memory means it can at least load the enormous 64GB and 128GB models. This article lays out, as a buyer’s guide, which models are viable on a MacBook or MacBook Air by memory tier, how the Air and Pro differ, and how it all stacks up against an RTX 5090.

The MacBook’s Quirk — ‘Unified Memory’ Is Your VRAM

On a regular PC, the CPU’s RAM and the GPU’s VRAM are separate. A local LLM only runs fast once it’s loaded into the GPU’s VRAM, so the graphics card’s VRAM (typically 8–32GB) becomes your ceiling. An Apple Silicon Mac, by contrast, shares a single pool of “Unified Memory” between the CPU and GPU. In other words, a Mac’s RAM is its VRAM, so on a 64GB Mac you can in principle devote most of that pool to the model.

This is the Mac’s decisive advantage for local AI. Because you can secure a lot of memory relatively cheaply, it becomes possible for an individual to run a 70B-class large model on a single laptop. Two caveats come attached, though. First, macOS reserves a chunk for the system, so the amount you can actually use is roughly 70% of the unified memory (an 8GB Mac really only has 5–6GB free for a model). Second, capacity is one thing, but memory bandwidth (speed) varies enormously by chip tier, so the same model can feel wildly different in real-world speed (more on this below).

Photo by Maxim Hopman on Unsplash

What Runs at Each Memory Tier — At a Glance

Here’s which models are practical at each unified-memory tier, based on Q4 quantization.

To hit only the essentials — 8GB is tight even as a local-LLM starter kit. Small 3–4B models work, but even a 7–8B model strains it. If you’re buying new today, 16GB is the bare-minimum line (recent MacBook Airs start at 16GB anyway), and if you’re serious about it, 32GB is the value sweet spot (32B-class recent models run comfortably). If you’re aiming at 70B, you need 64GB or more.

Air vs Pro — It’s Not Just About Memory

Even at the same 16GB, a MacBook Air and a MacBook Pro give you a different local-LLM experience. There are two reasons.

• Memory bandwidth: How fast the model can be read in governs generation speed (tok/s). The base M-series chip (Air) has lower bandwidth, and it widens several-fold as you move to Pro, Max, and Ultra. In other words, the same model runs far faster on a MacBook Pro (Max) than on an Air. Even with plenty of capacity, a narrow bandwidth makes large models feel sluggish.
• Heat (throttling): The MacBook Air has no fan. Work that hammers the GPU for a long stretch, like an LLM, builds up heat and gradually shaves off performance (throttling). It’s fine for brief use, but for long generations or sustained work, the fan-equipped MacBook Pro is more stable.

To sum up — if you’ll run small models lightly and on the go, a MacBook Air (16–24GB) is plenty, and if you’ll run 32B-and-up often and long and speed matters, a MacBook Pro (Pro/Max, 32–64GB+) is the right call.

Photo by Giorgio Trovato on Unsplash

Hands-On — Running Ollama on a MacBook

The easiest way to run a local LLM on a Mac is Ollama. Install it and it automatically detects and uses Apple Silicon’s Metal backend, so there’s no separate GPU setup to do.

# 16GB MacBook — 8~14B recommended
ollama run llama3.1:8b

# 32GB MacBook — 32B-class sweet spot
ollama run qwen2.5:32b

Two practical tips. First, watch memory pressure. When the model plus context exceeds your usable memory, macOS starts swapping and slows to a crawl. If memory pressure is yellow or red in Activity Monitor, step down a tier to a smaller model or stronger quantization. Second, context length eats memory too — the longer the document you feed in, the more headroom you need, so it’s safest to keep one tier of margin below the ceilings in the chart above.

MacBook vs RTX 5090 — Which Should You Pick

If you’re weighing a MacBook against a gaming GPU for local AI, they have opposite personalities.

In one line — “large models, quietly and on the move” is the MacBook; “small-to-mid models at maximum speed” is the 5090. The detailed 5090 breakdown is in How Many Billion Parameters an RTX 5090 Can Run.

So What — How Should You Buy a MacBook for Local AI

Compress the buying criteria into one sentence — “generous on capacity for the future, matched on tier to your use case.” If you plan to use local LLMs seriously, I recommend 16GB minimum, 32GB if you can. You can’t add memory later (Apple Silicon is soldered on-chip), so bumping one tier at purchase pays off for years. If speed also matters, go for a MacBook Pro’s Pro/Max chip; if you’ll use it lightly, a MacBook Air at 16–24GB is plenty.

Above all — before the hardware numbers, first decide “what size model is enough for my work.” Practical tasks like summarizing, coding, and document Q&A are usually well served by 14–32B, and if so, there’s no need to chase a top-end 128GB model. For most individuals, a 32GB MacBook is the realistic sweet spot for local AI.

Local LLM Series

• Part 1 — Getting Started with Local LLMs
• Part 2 — Running a Local LLM on Your iPhone
• Hermes AI × LM Studio Local Integration
• How Many Billion Parameters an RTX 5090 Can Run

Sources

• SitePoint — Local LLMs on Apple Silicon (2026) (models by memory tier)
• InsiderLLM — Best Local LLMs for Mac 2026 (M1–M5 benchmarks)
• ModelFit — Best LLM for MacBook by RAM tier (recommendations by RAM tier)