Intermediate 11 minMacBook Pro

Which LLM on MacBook Pro M1 Pro / Max (16–64 GB) ?

Released in late 2021, the MacBook Pro M1 Pro / Max remains an outstanding machine for local AI in 2026, especially in the Max 64 GB version. 200 to 400 GB/s bandwidth, active fans (no throttling), and up to 64 GB of unified memory: a MoE Qwen 3.6 35B runs at ~34 tokens/second, and a 30B model fits in full-quality Q8—something impossible on an equivalent mainstream laptop. This guide explains how to make the most of this machine today.

Choosing a machine? Our picks by budget →

By Mohamed Meguedmi·Update 2026-08-27·Tested on macOS 14+
Recommended hardware

Buying alternative for this guide: GMKtec EVO-X2 64GB / 1TB (Ryzen AI Max+ 395).

A mini PC is a complete machine: check the available memory and engine compatibility. It does not replace macOS/MLX or CUDA.

Why this choice? Our complete guide on GMKtec EVO-X2 64GB / 1TB (Ryzen AI Max+ 395) →

Compare all options by budget, from €800 to €3,500 →

On the go: which laptop for local AI →

Affiliate links — possible commission at no extra cost to you. As an Amazon Associate, BestLLMfor earns from qualifying purchases.

#M1 Pro / Max MBP in 2026

Output
October 2021. Still supported by macOS Sequoia (15) in 2026.
M1 Pro
8 or 10 CPU cores + 14 or 16 GPU cores, 200 GB/s bandwidth, 16 or 32 GB of RAM.
M1 Max
10 CPU cores + 24 or 32 GPU cores, 400 GB/s bandwidth, 32 or 64 GB of RAM.
Cooling
Fans active—no throttling during LLM use. Can run continuously for 24 hours.
2026 resale value
M1 Pro and M1 Max MBPs: used, with the price varying by condition.
→
Why it is a bargain in 2026
A used 64 GB M1 Max MBP (variable price) runs the best 2026 MoE models (Qwen 3.6 35B, Qwen3-Coder 30B) at full-quality Q8. For the same LLM capacity on PC, you need 2 × RTX 3090 (used, variable price) plus a compatible CPU / motherboard / PSU (~€800), for a significant total that is noisy and power-hungry.

#1. M1 Pro vs M1 Max: choosing

M1 Pro (200 GB/s)
Excellent for 8–12B models. A 30–35B MoE (Qwen 3.6 35B-A3B) fits on the 32 GB version, but not on 16 GB.
M1 Max (400 GB/s)
2× the bandwidth = 2× the inference speed on models ≥ 13B. Essential for running a 30B in full-quality Q8.
GPU cores
The 32-core M1 Max is 15–20% faster than the 24-core version on 8B models. The gap widens with 27–35B models.

#2. 16, 32, 64 GB: which models

Models recommended for an MBP M1 configuration
ModelQuantModel RAMM1 Pro 16 GBM1 Pro 32 GBM1 Max 64 GB
Qwen 3.5 9BQ5_K_M7 GB283143
Granite 4.2 8BQ5_K_M6 GB303447
Gemma 4 12BQ5_K_M9 GB182130
Qwen 3.8 27BQ4_K_M18 GB—1118
Mistral Small 24BQ5_K_M16 GB—1018
Qwen 3.6 35B-A3B (MoE)Q4_K_M23 GB—2234
Qwen3-Coder 30B-A3BQ8_032 GB——24

#3. Tokens/sec benchmarks

MBP M1 Max 32-core GPU, 64 GB, Ollama, wall power — ballpark figures
ModelQ4_K_MQ5_K_MQ8_0
Qwen 3.5 9B46 t/s43 t/s28 t/s
Granite 4.2 8B50 t/s47 t/s30 t/s
Gemma 4 12B32 t/s30 t/s19 t/s
Qwen 3.8 27B18 t/s16 t/s10 t/s
Qwen 3.6 35B-A3B34 t/s—22 t/s

#4. Installation and setup

Terminal — complete setup
# Homebrew
/bin/bash -c "$(curl -fsSL https://raw.githubusercontent.com/Homebrew/install/HEAD/install.sh)"

# Ollama + démarrage auto
brew install --cask ollama
brew services start ollama

# Modèles conseillés
ollama run qwen3.5:9b        # M1 Pro 16 Go
ollama run qwen3.8:27b       # M1 Pro 32 Go
ollama run qwen3.6:35b       # M1 Max 64 Go (MoE 35B)

#5. Determine the GPU memory limit

On an MBP M1 Max with 64 GB, macOS allocates ~43 GB to the GPU by default. To load a 30-35B model in Q8 (~35 GB) with an extended context—the KV cache for a 256k context can exceed 10 GB on its own—the total quickly surpasses 43 GB, so you need to raise the limit.

The Mac Kit

Your MacBook Pro M1 can do more than you think. The Mac kit shows you how to push past the GPU memory limit (ch. 2), choose the right model for your chip (ch. 4), and fix what breaks: Metal, memory, swap (ch. 14).

  • Lifetime online access
  • PDF + files
  • Lifetime updates
The Mac Kit →or all kits for life — $49 →
Increase allocatable GPU memory
# Vérifier la valeur actuelle
sysctl iogpu.wired_limit_mb

# Relever à 56 Go (64 Go - 8 Go système)
sudo sysctl iogpu.wired_limit_mb=57344

# Rendre permanent
echo "iogpu.wired_limit_mb=57344" | sudo tee -a /etc/sysctl.conf
!
Leave ≥ 8 GB for macOS
Below that point, the system starts swapping and inference slows dramatically. 8 GB is the minimum headroom on a 64 GB MBP; stay at 6 GB only if you close all Electron apps and do not use a browser.

#6. Running a large 2026 model on M1 Max

The 64 GB M1 Max MacBook Pro runs the largest 2026 models at full quality: a 30-35B MoE in Q8, or several models loaded in parallel. Here's the optimal configuration:

MoE Qwen 3.6 35B on M1 Max 64 GB
# Augmenter la RAM GPU (fait une seule fois)
sudo sysctl iogpu.wired_limit_mb=57344

# Activer flash attention + KV cache quantifié
export OLLAMA_FLASH_ATTENTION=1
export OLLAMA_KV_CACHE_TYPE=q8_0
brew services restart ollama

# Lancer le modèle
ollama run qwen3.6:35b
# Comptez ~34 tokens/seconde (MoE, 3B actifs) — contexte étendu confortable
i
Battery life with large models
About 1h30 of battery life during continuous chat on a 35B MoE (consumption ~40 W). For long sessions, plug it in—the MBP M1 Max heavily throttles the GPU on low battery.

#Upgrade to M3 Max or M4 Max?

M3 Max 16-core (2023)
+50% raw speed vs. M1 Max. 128 GB RAM possible (vs. 64 GB). Justifies the upgrade if you want to push a 30–35B MoE beyond 50 tok/s.
M4 Max 16-core (2024)
+90% vs. M1 Max, 546 GB/s bandwidth. 128 GB RAM. Best Mac laptop for LLMs in 2026.
Stick with the M1 Max 64 GB
Entirely viable: 35B MoE at ~34 tok/s, 27B dense at 18 tok/s. No reason to upgrade before 2027 if you're satisfied with the current speed.

#Frequently asked questions

MBP M1 Pro 16 GB or M1 Max 32 GB for local AI?+
The M1 Max 32 GB for a real improvement: double the bandwidth (400 vs 200 GB/s) means double the speed on 13B+ models, plus 20 GB of additional usable memory. However, if you only plan to run 8-9B models, the M1 Pro 16 GB is more than sufficient.
Can you run a large model on a 32 GB MBP M1 Max?+
Yes: a 30–35B MoE in Q4 (~23 GB, like Qwen 3.6 35B-A3B) fits in 32 GB and runs quickly thanks to its 3B active parameters. By contrast, a 30B model in Q8 (~32 GB) or a dense model ≥ 27B in Q8 saturates 32 GB, so move to 64 GB.
Is the M1 Pro MBP better than a RTX 3060 PC for LLMs?+
For Qwen 3.5 9B Q4: comparable speed (30-45 tok/s). Mac advantage: 16 GB usable vs. 12 GB VRAM, quiet operation, no drivers. PC advantage: the CUDA ecosystem and simpler fine-tuning.
Do the fans get loud during LLM inference?+
Audible on a large model in continuous chat (RPM ~4000, equivalent to a 2019 Intel MacBook Pro compiling). Quiet on 8-12B (RPM ~2500). Much quieter than a gaming PC laptop under the same GPU load.
Do LLMs need a 16" or 14" screen?+
No impact on performance. The 16" has better heat dissipation (larger chassis), so it sustains the load slightly longer. Choose 14" if portability is the priority, 16" if long sessions take precedence.
Ollama or MLX on an M1 Max MBP?+
Ollama remains the foundation for 95% of use cases (simple, robust, OpenAI API). MLX is useful for fine-tuning an 8–9B model locally or for MLX-native models that benefit from the M1 Max. Both coexist without conflict.
Did this guide help you?

Feedback, an error, or a clarification? Let us know—it improves the guide for everyone.