Which LLM on Mac Studio (M2 / M3 / M4 Ultra, 64–512 GB) ?

#Mac Studio in 2026

Mac Studio M2 Ultra (2023)

24 CPU cores + 60 or 76 GPU cores, 800 GB/s, 64/128/192 GB RAM. Still readily available used starting at €2,800.

Mac Studio M4 Max (2025)

16 CPU cores + 40 GPU cores, 546 GB/s, 36–128 GB of RAM. Replaces the M2 Max. No longer sold new by Apple: replaced in late September 2026 by the M5 Max (starting at €2,999).

Mac Studio M3 Ultra (2025)

28 CPU cores + 60 or 80 GPU cores, 800 GB/s, 96/256/512 GB RAM. The monster. Sold for 4499 € (96 GB) to 11 299 € (512 GB) until the M5 Ultra arrived; no longer in the Apple catalog at the end of September 2026 (used or remaining stock).

Power consumption

15 W idle, 70-130 W during inference depending on the configuration. The 512 GB M3 Ultra peaks at ~150 W with 405B.

#1. M2 Ultra vs M3 Ultra vs M4 Max

M4 Max

The entry-level Mac Studio. 546 GB/s, 128 GB maximum. Better than an M4 Max MBP: no throttling, lower cost (desktop enclosure).

M2 Ultra

First-generation Ultra. 800 GB/s, up to 192 GB. Used price: €2,800–4,500. Still competitive for 70B–123B.

M3 Ultra

Latest Ultra (2025). The same 800 GB/s as the M2 Ultra, but with M3 architecture (+20% in practice). 96 to 512 GB of RAM. For anyone looking to go beyond 200B.

#2. How far should you push the RAM?

96 GB (base M3 Ultra)

€4,499 for the M3 Ultra in the catalog, no longer sold new (M5 Ultra 96 GB: €6,599). Runs gpt-oss 120B (MXFP4, 63 GB) at ~50 tok/s and ~30B MoEs (Qwen 3.6 35B-A3B) at ~70 tok/s, with room for a long context. Sweet spot for large MoEs up to ~120B.

256 GB (M3 Ultra + 256)

€7299 listed for the M3 Ultra, no longer sold new. Runs Llama 4 Maverick 402B Q4 (230 GB) at ~34 tok/s and DeepSeek V3 671B Q2 (~240 GB). For anyone who wants 400B+.

512 GB (M3 Ultra max)

€11,299 at the time of the M3 Ultra’s listing, no longer sold new (M5 Ultra 512 GB: late October 2026). DeepSeek V3 671B Q4 (380 GB) works, as does Qwen3-235B-A22B in Q8, with 128k+ context on large MoE models. Research/R&D territory.

192 GB (M2 Ultra max)

Runs Qwen3-235B-A22B Q4 (133 GB) and Maverick 402B in Q3. Functionally equivalent to the 256 GB M3 Ultra for a little less on the used market. A good compromise.

#3. Ultra-exclusive models

#4. 120B, 235B, and 671B benchmarks

#5. LLM workstation setup

# Ollama + optimisations max
brew install --cask ollama

launchctl setenv OLLAMA_FLASH_ATTENTION 1
launchctl setenv OLLAMA_KV_CACHE_TYPE q8_0
launchctl setenv OLLAMA_NUM_PARALLEL 4
launchctl setenv OLLAMA_HOST "0.0.0.0:11434"

# Relever mémoire GPU à 480 Go (sur 512 Go)
sudo sysctl iogpu.wired_limit_mb=491520
echo "iogpu.wired_limit_mb=491520" | sudo tee -a /etc/sysctl.conf

brew services start ollama

# Pour DeepSeek V3 671B : passer par llama.cpp maison
git clone https://github.com/ggerganov/llama.cpp && cd llama.cpp
make GGML_METAL=1 LLAMA_METAL_EMBED_LIBRARY=1

# Lancement
./build/bin/llama-server \
  -m ./models/DeepSeek-V3-671B-Q4_K_M.gguf \
  -ngl 99 -fa -c 16384 \
  -ctk q8_0 -ctv q8_0 \
  --host 0.0.0.0 --port 8080

#6. DeepSeek V3 671B on M3 Ultra 512 GB

This is the flagship use case for the 512 GB Mac Studio. The model is 380 GB in Q4_K_M, fits in RAM on the 512 GB configuration, and runs at 16k context at ~16 tok/s (MoE, ~37B active parameters out of 671B). Quality is close to the best proprietary models for most tasks.

Download

About 380 GB via Hugging Face. Allow 4-8 hours depending on the connection.

Load time

~60 seconds the first time (SSD → RAM). Instant afterward as long as the model remains in memory.

Inference usage

~150 W. Reaches 70-75 °C. Fan is audible but not loud (far quieter than an equivalent PC).

When it is worth it

R&D, teams that want a 100% offline frontier model, internal benchmarks, sensitive datasets. For production, the cloud API is more economical.

#Mac Studio vs. PC 2× RTX 5090

2026 price

Mac Studio M3 Ultra 512 GB: ~€11,300 at list price through September 2026; no longer sold new (M5 Ultra 512 GB announced for late October). PC 2× RTX 5090 + platform: ≈ €15,000.

Useful memory for LLMs

Mac: 512 GB unified. PC: 2 × 32 GB VRAM = 64 GB. For 400B+ (Maverick 402B, DeepSeek 671B), the PC simply cannot load the model.

Speed on an ~30B (Qwen 3.6 35B-A3B)

Dual-5090 PC: ~120 tok/s (CUDA, fits in VRAM). M3 Ultra Mac: ~70 tok/s. PC advantage for what fits in VRAM.

Speed on DeepSeek 671B / Maverick 402B

Mac: 16-34 tok/s (MoE). PC: impossible without massive CPU offloading (< 2 tok/s).

Power consumption

Mac Studio: 150 W peak. PC with 2× 5090s: 900 W max. Expect the PC electricity bill to be 3× higher.

Noise

Mac: slightly audible. PC with 2× 5090s: very loud (GPU fans at full speed).

Verdict

The Mac Studio wins on capacity (200B+ models) and efficiency. The PC wins on raw speed for models that fit in VRAM (≤ ~30 GB). Choose based on your target model profile.

#Frequently asked questions

Prices change quickly: every Monday and Thursday, our tracker records the lowest price for local AI graphics cards, along with the price per GB of VRAM.

• Local AI graphics card prices (updated twice a week)