How Much RAM for Local AI on a Mac? 16GB to 512GB (2026)

What each RAM tier runs

Each pick is the index’s best-ranked model that fits the tier, reaches 15 tok/s on the tier’s reference Mac and has a standard Mac runtime (the rule). Picks are as of 24 October 2026; the live version, which updates with every build, is Best LLM by Mac RAM.

“Needs” is the index’s minimum RAM for the model. The 16 GB Mac mini M6 is a binned part (153 GB/s, against 170 GB/s for the 24 and 32 GB models), so its figures are always estimates. For the 16 GB tier in detail, see the best local LLM for a 16 GB Mac.

How much memory a model needs

Three things share your unified memory: the model’s weights, its key-value (KV) cache, which grows with the length of the conversation, and macOS with your other apps. The index’s minimum-RAM rule takes the smallest build of at least 3 bits that has a working Mac runtime and adds about 2 GB. At 4-bit (Q4_K_M) that is roughly 0.57 GB per billion parameters:

• A 12B dense model is about 7 GB at 4-bit, comfortable on 16 GB.
• A 27B dense model is about 16–17 GB; Qwen3.8-27B’s index minimum is 19 GB.
• A 35B mixture-of-experts model is about 20 GB; Qwen 3.6-35B-A3B’s minimum is 24 GB.
• A 70B dense model is about 40 GB. tbreak’s Mac Studio M5 Max review recorded a 41.2 GB peak for a 4-bit Llama 3.3 70B; the index’s minimum is 44 GB, so a 48 GB Mac is the floor and 64 GB is comfortable.

The older “1.5× the file size” rule of thumb is too generous for small models and too tight for long contexts. The real limit is the GPU’s share: macOS does not let the GPU wire all of unified memory. Metal reports the cap as recommendedMaxWorkingSetSize, and llama.cpp prints it when a model loads. If a model fits in RAM but loads slowly or falls back to the CPU, the mlx-lm README suggests raising the wired-memory limit on macOS 15 or newer with sudo sysctl iogpu.wired_limit_mb=N, where N is larger than the model in megabytes and smaller than your RAM. It resets at restart; leave a few gigabytes for macOS.

When a model does not fit, macOS swaps to the SSD and decode slows to a crawl. Pick a smaller quantization (see which quantization to use) or a smaller model rather than relying on swap.

Mixture-of-experts: RAM sets what fits, active parameters set speed

A mixture-of-experts (MoE) model splits most of its layers into many small “experts” and routes each token through only a few. Qwen 3.6-35B-A3B has 35B parameters in 256 experts, but each token uses 8 routed experts plus 1 shared one, about 3B parameters (model card). All 35B must sit in memory, which is why it needs a 24 GB Mac. Only about 3B are read for each token, which is why it is fast.

On a Mac mini M6 (170 GB/s), that MoE model runs at community64 tok/s (32 GB) while Alibaba’s dense Qwen3.8-27B runs at community9 tok/s (24 GB), both community-measured. The dense model is the more capable of the two on the index’s capability score (46 against 38 out of 50), but it reads all 27.8B parameters for every token. So an MoE model does not save RAM; it turns the RAM you have into speed. More in MoE vs dense LLMs.

Same RAM, different speed

Two Macs with the same memory run the same models at different speeds, because decode is limited by memory bandwidth. A 24 GB Mac mini M6 (170 GB/s) and a 24 GB Mac mini M5 Pro (307 GB/s) both fit Qwen 3.6-35B-A3B, at community64 tok/s and community79 tok/s respectively. For a dense model the gap tracks the bandwidth ratio more closely: Qwen3.8-27B at community9 tok/s on the M6 against sourced17 tok/s on the M5 Pro. Decide the RAM first, then pay for bandwidth if the models you want are dense or large. The Apple Silicon chip comparison lists every chip’s bandwidth.

Which tier to buy

Unified memory is part of the chip package, so the tier you buy is the tier you keep. In order of how often it is the right answer:

• 32 GB for most people who want local AI to be useful every day. It holds the index’s #1, Qwen 3.6-35B-A3B, with room for long conversations and a browser, and it fits the strongest dense model, Qwen3.8-27B.
• 24 GB is the minimum for Qwen 3.6-35B-A3B. It works, but a long context and a few open apps will push it.
• 16 GB runs 12B-class models and GPT-oss 20B well enough for chat, summaries and light coding.
• 48–64 GB if you want the 80–120B mixture-of-experts models (Qwen3-Coder-Next, GPT-oss 120B) or several models loaded at once. At 64 GB the index’s pick is the faster LFM2.5-24B-A2B.
• 128 GB (an M3, M4 or M5 Max) for the 100B-plus tier at 4-bit and 280B-class models at 2-bit.
• 192–512 GB (an Ultra chip in a Mac Studio) for MiMo-V2.6-Flash, GLM 5.2 and, at 512 GB, the 1T-parameter models at about 3-bit.

The Mac Advisor turns a budget and a use case into an exact configuration, and every Mac has its own ranking under Best LLM by Mac.

Why unified memory matters

On a PC, a model runs fastest when it fits in the graphics card’s own memory: 24 GB on an RTX 4090. Anything larger spills into system RAM over the PCIe bus and slows down sharply. On Apple Silicon the CPU and GPU share one pool of unified memory, so the GPU can address most of a 64 or 128 GB Mac directly. That is why a Mac can run models no single consumer graphics card can hold, even though a high-end graphics card’s own memory is faster.

Frequently Asked Questions