Which LLM on a Mac mini M4 / M4 Pro (16–64 GB) ?

#Mac mini M4 and local LLMs: what the machine can do

A Mac mini makes a good local LLM server because its unified memory is shared between the CPU and GPU, with no separate VRAM. Two figures determine what you can do with it. Capacity (16 to 64 GB depending on the chip) determines the model size that can load: 7 to 12 billion parameters on 16 GB, and 24 to 35 billion on 32 to 48 GB. Bandwidth determines speed: 120 GB/s for the M4 chip and 273 GB/s for the M4 Pro, according to Apple. An LLM rereads its weights for every generated token, so maximum throughput is roughly the bandwidth divided by the weight size. For a 7B model in Q4_0, llama.cpp's public table reports 24.1 tokens/s for an M4 and about 50 for an M4 Pro. Apple introduced Mac mini M6 and M5 Pro models in August 2026 that replace this lineup; the M4 was still available from some retailers in late August.

Mac mini M4

10-core CPU, 10-core GPU, 120 GB/s, 16 GB unified memory (24 or 32 GB optional), Thunderbolt 4.

Mac mini M4 Pro

12-core CPU, 16-core GPU (14 and 20 optional), 273 GB/s, 24 GB of memory (48 or 64 GB optional), Thunderbolt 5.

Format and noise

12.7 × 12.7 × 5.0 cm, 0.67 kg (M4) or 0.73 kg (M4 Pro); 5 dBA at idle, according to Apple.

Advertised consumption

M4: 4 W at idle, 65 W at maximum. M4 Pro: 5 W and 140 W. Apple does not publish an inference figure.

#Mac mini M4 or M4 Pro: bandwidth makes the difference

Between the two chips, the real difference is not maximum memory (32 GB versus 64 GB) but bandwidth, multiplied by 2.3 (273 ÷ 120). In the llama.cpp table, the measured ratio is 2.1 for generation (50.7 ÷ 24.1 in Q4_0) and 2.0 for prompt processing (440 ÷ 221).

The M4 is enough if

you are on your own, target models with up to 12 billion parameters (Qwen 3.5 9B, Gemma 4 12B) and accept around twenty tokens/s at best. Choose 24 GB to keep an embeddings model loaded as well.

The M4 Pro is the clear choice if

you are targeting models with 14 to 27 billion parameters, a 30- to 35-billion-parameter MoE (48 GB minimum), a coding agent, or multiple users. Beyond 64 GB, you need a Mac Studio.

#VRAM on Mac mini: unified memory and its GPU ceiling

A Mac mini has no dedicated VRAM, but the GPU cannot use all the memory. According to llama.cpp users, the default ceiling is two-thirds to three-quarters of the RAM; at launch, llama.cpp displays the recommendedMaxWorkingSetSize line, which gives the actual value. The iogpu.wired_limit_mb setting defaults to 0, meaning “system default policy,” not “unlimited.”

The ceiling must hold the weights, the KV cache (the context memory, which grows with the conversation), and compute buffers. To load a model that is close to the limit, raise the ceiling: MLX recommends a value higher than the model size in megabytes but lower than the machine’s memory. Leave several gigabytes for macOS, or speed will collapse. A manual sysctl setting may not survive a reboot.

# Valeur actuelle : 0 = plafond par défaut du système
sysctl iogpu.wired_limit_mb

# 40 Go pour le GPU sur 48 Go de RAM (40 x 1024 = 40960 Mo)
sudo sysctl iogpu.wired_limit_mb=40960

#Which model for which memory: what fits in a Mac mini

The table compares the file sizes published by Ollama with a cap of two-thirds of RAM (a conservative assumption). “Yes”: the files use less than 70% of the cap, leaving room for the KV cache. “Barely”: they fit, with a short context. “No”: they exceed it. These are calculations, not measurements.

#Mac mini M4 16 GB: models to prioritize

On 16 GB, the ceiling is around 11 GB. Qwen 3.5 9B (6.6 GB at Ollama) leaves about 4 GB for the KV cache and buffers. Gemma 4 12B (7.6 GB) works with a short context. A 14 GB model such as Mistral Small 24B overflows: Ollama splits it between the GPU and CPU, and speed drops. The ollama ps command displays the split; aim for 100% GPU.

#Mac mini speed: theoretical ceiling and published measurements

Generation throughput is limited by bandwidth: tokens per second ≈ bandwidth (GB/s) ÷ weights read per token (GB). This calculation gives an upper bound, never a measurement. For a dense 14 GB model, 120 ÷ 14 gives 8.6 tokens/s at best on an M4, and 273 ÷ 14 gives 19.5 on an M4 Pro.

The published measurements remain below this ceiling. Here are the figures from the table that the llama.cpp community maintains for the Apple chips, with Llama 2 7B. These are not our measurements, and they cover an older model in Q4_0 and Q8_0, but they show the performance ratio between the chips.

Moving from Q8_0 to Q4_0 speeds up generation by a factor of 1.8 on M4 and 1.7 on M4 Pro: file size matters more than computation. GPU cores matter mainly for prompt processing (364 versus 440 tokens/s on M4 Pro). The table contains no M6 measurements as of the date consulted.

#Mac mini M6 and M5 Pro: what the new lineup changes for an LLM

Apple announces up to 4.8 times faster prompt processing on the Mac mini M6 in LM Studio than on the M4, and up to 4 times faster on the M5 Pro than on the M4 Pro. These gains apply to prompt processing, which benefits from the neural accelerators added to each GPU core. Generation remains governed by memory bandwidth and improves much less: from 12% (307 versus 273 GB/s) to 42% (170 versus 120 GB/s).

A base M6 at 153 GB/s has only 28% more bandwidth than the M4: the theoretical ceiling for a 9B rises from about 18 to 23 tokens/s. The new lineup is mainly beneficial for long-context use cases: RAG that injects long documents, or a coding agent that rereads a repository.

Nine, 30- to 35-billion-parameter model

An M5 Pro with at least 48 GB, or 64 GB to work comfortably with a 35-billion-parameter MoE: the M6 tops out at 32 GB.

M4 still in stock

A reasonable choice for 7 to 12 billion parameters if the price gap with the M6 is significant. This stock is limited.

#Install an LLM server on a Mac mini with Ollama

Ollama on macOS works like an application, with its settings configured through environment variables defined with launchctl. By default, Ollama listens only on 127.0.0.1, port 11434: without changes, no other device can use it.

1. 01
   Install the Ollama app

   Download Ollama from ollama.com and place the app in Applications. On first launch, it offers to create the ollama command link.
2. 02
   Open the port to the local network

   Set OLLAMA_HOST with launchctl (command below), then restart the application. The address 0.0.0.0 opens the API to the entire network, without authentication: limit it to a trusted network.
3. 03
   Prevent sleep

   In System Settings, under Energy, enable the option that prevents automatic sleep when the display is off; without it, the service may stop.
4. 04
   Download a suitable model

   In the table, choose a model marked “yes” for your memory, such as qwen3.5:9b on 16 GB.
5. 05
   Check from another device

   Send a request to the Mac mini’s address (.local name or IP), then run ollama ps: the Processor column should show 100% GPU.

# Écoute sur tout le réseau local, puis relancer l'application Ollama
launchctl setenv OLLAMA_HOST "0.0.0.0:11434"

# Modèle pour 16 Go de mémoire
ollama pull qwen3.5:9b

# Test depuis un autre appareil (remplacez l'adresse)
curl http://mac-mini.local:11434/api/chat -d '{"model": "qwen3.5:9b", "messages": [{"role": "user", "content": "Bonjour"}], "stream": false}'

Two settings often copied from elsewhere are useless here: Ollama's documentation says Flash Attention is enabled automatically when the hardware allows it, and OLLAMA_ORIGINS concerns browser extensions. By contrast, quantizing the KV cache (OLLAMA_KV_CACHE_TYPE=q8_0) cuts its size roughly in half for a very small loss, which matters on 16 GB when the context grows.

#LM Studio: the alternative to Ollama

LM Studio requires an Apple Silicon chip, macOS 14 or later, and 16 GB of recommended RAM. It loads GGUF models like MLX, starting with version 0.3.4. Without a display, its llmster daemon starts with lms daemon up. The “Serve on Local Network” option exposes the API to the network; token authentication is not enabled by default. Ollama’s MLX engine, presented in preview in March 2026, required more than 32 GB: an M4 16 GB did not benefit from it.

#What a Mac mini can serve: chat, RAG, agents, cluster

Self-hosted chat server

The Ollama and LM Studio APIs accept OpenAI-format clients (Open WebUI, Zed, scripts). Placed near the router, the Mac mini serves the whole house.

Document RAG

You need to load both the generation model and an embeddings model. nomic-embed-text weighs 274 MB: on 16 GB, it fits alongside Qwen 3.5 9B.

Code agent

An agent sends long prompts at every step: prompt reading matters more than generation, and that's where M6 and M5 Pro gain the most.

#How many concurrent users?

By default, Ollama handles one request at a time per model (OLLAMA_NUM_PARALLEL is 1). Each parallel request increases the allocated context: according to the documentation, 4 requests with a context of 2,000 tokens allocate 8,000 tokens. Cache memory = cache for one conversation × parallel requests. Each user gets a slower response; no reliable measurement is available here for a specific number of users.

#Link multiple Mac minis into a cluster

Only the M4 Pro offers Thunderbolt 5 (120 Gb/s); the M4 is limited to Thunderbolt 4 (40 Gb/s). The exo project supports RDMA over Thunderbolt 5, which would reduce latency between machines by 99%, and says it speeds up models as devices are added, on macOS 26.2 or later. Its benchmarks cover Mac Studio, not Mac mini. A cluster is primarily used to load a model that is too large for one machine.

• Share Ollama on your local network (family, team)
• Securing your Ollama server: authentication and reverse proxy
• Exo: turn multiple machines into a home LLM cluster
• Optimize your Mac Apple Silicon: GPU ceiling, Flash Attention, KV cache
• Hardware profile: Mac mini M6
• Hardware profile: Mac mini M5 Pro

#Mac mini M4 or MacBook Air M4: same chip, different use case

The 13-inch M4 MacBook Air uses the same chip, with the same memory bandwidth (120 GB/s) and the same 16, 24, or 32 GB options: the speed ceiling is identical. What changes is the chassis. The previous version of this guide estimated a 20% slowdown for the Air after ten minutes; no source confirms this, so it has been removed. For a service running continuously, the Mac mini is the natural choice; for a mobile workstation (1.24 kg, including the display and battery), the MacBook.

• Source: Apple, Mac mini (2024) technical specifications
• Source: Apple, announcement of the M6 and M5 Pro Mac mini (August 2026)
• Source: llama.cpp, performance on Apple M-series chips
• Source: Ollama documentation, FAQ (environment variables, KV cache)
• Source: LM Studio documentation, server mode without a UI

#Frequently asked questions