Can an Apple M4 Max 128GB run Llama 3.3 70B Instruct?

In practice

Llama 3.3 70B Instruct sits comfortably on Apple M4 Max 128GB. At IQ4_XS it takes 37.92 GB of the 96 GB available, which leaves 58.08 GB spare, enough for the 8k window we assumed and for a window of up to about 128k tokens before memory runs out. Generation runs at about 10.1 tokens per second, faster than most people read, so a 500-token answer takes about 50 seconds.

Of the memory in use, 92% is the weights and 7% is the attention cache. With this much room the question is quality, not fit. The next step up, Q4_K_M, also fits at 42.68 GB, but runs at 8.9 tokens per second, so IQ4_XS is where we would stay.

Reading a 4,000-token prompt, roughly a long email or a few pages of text, takes about 2 minutes before the first word appears. That part is limited by compute and not by bandwidth, so it follows a different spec from the writing speed.

The numbers

Llama 3.3 70B Instruct has 70.55 billion parameters across 80 layers. The Apple M4 Max 128GB has 128 GB of unified memory at 546 GB/s, of which about 96 GB is left for a model once the operating system has taken its share.

9 of the 9 levels fit entirely. Levels marked degraded are listed for completeness, not as advice.

How long a conversation it holds

Memory rises with the length of the conversation because the attention cache keeps a key and a value for every token. At IQ4_XS, the longest window that still fits on this device is 128k tokens. Setting the window to what you need and no more is the cheapest way to buy back memory.

The settings we would use

llama-server -m <model>.gguf -ngl all -c 8192 -fa on

In Ollama, set OLLAMA_CONTEXT_LENGTH to the window you want, run ollama ps while the model is loaded and check that the processor column reads 100% GPU. If it shows a split on a model that fits, the Ollama guide shows how to fix it. The llama.cpp flags guide explains each option above.

What else an Apple M4 Max 128GB runs well

For comparison, these are the other popular models that fit this device at a sound quantization, best pairings first.

The same model on nearby hardware

Memory decides whether Llama 3.3 70B Instruct loads and bandwidth decides how quickly it answers. Here is what a step down, a card with the same memory and a step up do with it.

Wait for the first word13 s

Then writes at10.1 tok/s

Whole answer30 s

YouWhy does my model use more memory when the conversation gets longer?

Model

Simulated from our estimate at a 512-token question, not a recording. Assumes nothing else is competing for the GPU.

Questions people ask

Keep going

Everything the Apple M4 Max 128GB runs, the full breakdown for Llama 3.3 70B Instruct, how the numbers are made in the methodology, or a different pairing in the calculator.