Which LLM on MacBook Pro M4 Pro / Max (24–128 GB) ?

#MacBook Pro M4 Pro and M4 Max: which configurations, and what bandwidth

On a MacBook Pro M4, the chip determines two things: bandwidth, which determines maximum generation speed, and maximum memory, which determines model size. According to the Apple technical specifications, the M4 Pro offers 273 GB/s, the M4 Max with 14 CPU cores and 32 GPU cores offers 410 GB/s, and the M4 Max with 16 CPU cores and 40 GPU cores offers 546 GB/s. Memory ranges from 24 GB to 128 GB depending on the chip. These are not interchangeable options: each chip opens up its own class of models.

Apple specifies that the M4 Max supports up to 128 GB of unified memory and that this allows developers to interact with language models of nearly 200 billion parameters. Apple specifies neither the quantization nor the context for this figure: do not read it as a promise for a dense Q4 model, since a dense 70B already weighs approximately 40 GB according to the site's benchmarks.

The specification describes the 14-inch MacBook Pro; the 16-inch model uses the same chips. The M4 chips have since been followed by an M5 generation: new units are no longer always available in the generation described here, and the used market determines its price. This page cites no prices; the /materiel-ia/macbook-pro-m4-max page provides up-to-date benchmarks.

#Pro, Max 14, or Max 16 cores: how to choose

1. 01
   Set your target model size first

   An 8- to 14-billion-parameter model runs on an M4 Pro with 24 GB. A 24B to 32B model requires 36 to 48 GB. A 70B requires at least 64 GB, with a short context.
2. 02
   Add context and a 25 to 30% margin

   The KV cache grows with context length, and macOS reserves some memory for itself. Multiply the model size by about 1.3 for a cautious estimate.
3. 03
   Check bandwidth against the desired speed

   Divide bandwidth by model size: that is the theoretical maximum speed. If it is too low for your use case, upgrade the chip rather than the memory.
4. 04
   Choose between a Max chip with 14 or 16 cores

   The 16-core model delivers 546 GB/s versus 410 GB/s, and, above all, access to 48, 64, and 128 GB: memory, as much as speed, justifies the difference.

This procedure explains why the M4 Pro with 48 GB and the 14-core M4 Max with 36 GB are not comparable: the former offers more memory, while the latter offers more bandwidth. For workloads where the model is small and responses need to be fast, the 14-core Max wins. To load a larger model, the 48 GB Pro wins, at the cost of slower generation.

#What bandwidth allows: theoretical ceilings

Each generated token rereads the active weights. Maximum speed is therefore bandwidth divided by weight in gigabytes, using the site's reference points: 14B about 9 GB, 32B about 19 to 20 GB, 70B about 40 GB in Q4. The table applies this formula to the three chips. These are calculated ceilings, never measurements: actual speed depends on the engine, context, and thermal load.

Mixture-of-experts (MoE) models are a different case: a 30B model with 3B active parameters takes about 19 GB of memory but reads only a few gigabytes of weights per token, so its ceiling is several times higher than that of a dense 32B model of the same weight. It’s the best deal for a MacBook Pro with 48 GB or more. Check actual throughput on /benchmarks or in third-party benchmarks, verifying the chip, quantization, and engine.

#MLX, Ollama, LM Studio: which engine on this machine

Ollama covers the essentials: one-click installation, local API, automatic GPU detection. MLX, Apple's framework, leverages the Mac's shared memory without copying between CPU and GPU; its mlx-lm package lets you generate text and fine-tune models on Apple silicon, with support for quantized models. LM Studio provides a graphical interface and MLX models. The comparative guide to MLX versus llama.cpp settles the performance question; there's no need to repeat its conclusions here.

1. 01
   Install Ollama

   Download it from ollama.com or with Homebrew. Automatic startup is described in the macOS installation guide.
2. 02
   Launch the selected model

   The ollama run t command downloads and then runs the model. Then check the GPU with ollama ps.
3. 03
   Configure the KV cache and flash attention

   Ollama automatically enables flash attention when the hardware supports it; the OLLAMA_KV_CACHE_TYPE variable quantizes the KV cache, with f16 as the default. It reduces the memory used by long contexts.
4. 04
   Increase the GPU memory limit if necessary

   The mlx-lm repository indicates that a model that fits in RAM can often be accelerated by raising the system's wired memory limit. This is an administrator-level system setting: follow the procedure in the macOS optimization guide, without copying a value found elsewhere.

brew install --cask ollama
ollama run qwen3:32b
ollama ps

# Long contexte : cache KV quantifié
export OLLAMA_FLASH_ATTENTION=1
export OLLAMA_KV_CACHE_TYPE=q8_0

#Long context: the use case where the Pro’s memory matters most

Long context is the main reason to choose 64 and 128 GB. Ollama sets a default context of 4 000 tokens under 24 GiB of memory, 32 000 between 24 and 48 GiB, and 256 000 starting at 48 GiB; its documentation recommends at least 64 000 tokens for web research, agents, and coding tools. A context of 64 000 tokens on a 32B model costs gigabytes of KV cache, in addition to the 19 to 20 GB of weights: on 48 GB, the headroom disappears quickly. A 64 GB or 128 GB machine leaves room for this context without requiring cache quantization.

• Quantizing the KV cache to save memory
• Understanding the context window

#MacBook Pro M4 versus a NVIDIA GPU, a Mac Studio, or an Air

The useful comparison is memory before speed. A RTX 4090 comes with 24 GB of GDDR6X, according to NVIDIA: beyond a model with 24 to 30 billion parameters, it must offload part of the model to system RAM, and speed drops. A MacBook Pro M4 Max with 64 or 128 GB can load a full 70B model. The card remains superior when the model fits in its memory and speed is the priority.

#M4 or M5: what the next generation changes

Apple has since released MacBook Pro models with M5 Pro and M5 Max chips. Their specifications list 307 GB/s for the M5 Pro, 460 GB/s for the M5 Max with 32 GPU cores, and 614 GB/s for the M5 Max with 40 GPU cores, versus 273, 410, and 546 GB/s for the M4 generation. Bandwidth increases by about 12 to 13%, so theoretical speed ceilings do too. The buying decision mainly comes down to equal memory: a used M4 Max with 128 GB opens up more models than a new M5 Pro with 48 GB.

#Limitations to know

Memory ceiling

Up to 128 GB. Beyond that, you need a Mac Studio or another platform.

No CUDA

Libraries that require CUDA do not run. MLX and llama.cpp cover inference, not the entire training ecosystem.

Fine-tuning

MLX supports low-rank and full fine-tuning with quantized models. On models with more than a few billion parameters, this remains slow and generates heat: reserve large workloads for a cloud service.

Battery

Continuous inference quickly drains the battery: plug in for long sessions.

• MacBook Air M4: 32 GB and 120 GB/s
• Mac Studio: the same chip on a desktop
• Mac mini M4 and M4 Pro
• Optimize a Mac Apple Silicon
• Hardware profile: MacBook Pro M4 Max

• Source: Apple technical specifications for the MacBook Pro M4 Pro and M4 Max
• Source: Apple technical specifications for the MacBook Pro M5 Pro and M5 Max
• Source: Apple, announcement of the M4 Pro and M4 Max chips
• Source: Ollama documentation, context
• Source: mlx-lm repository

#Frequently asked questions