The Mac Studio most people should buy: an 18-core M5 Max with a 32-core GPU (40 optional), more ports than any other Mac and room to cool it. The 36GB base model runs up to 30B-class 4-bit models; configured to 128GB with the 40-core GPU and 614GB/s it holds 70B-class models, and in llama.cpp's 7B 4-bit test it generated nearly 120 tokens a second against under 68 for the M5 Pro Mac mini. For hundreds of billions of parameters you need the M5 Ultra.
Against the line-up
Geekbench 6 CPU
Geekbench 6 CPU
Geekbench 6 GPU
Blender Benchmark
Geekbench AI
The whole Mac Studio line-up side by side, with prices
Every benchmark
Quoted from the publications that ran them; we don't test hardware ourselves. Figures from different publications were run on different systems and scenes, so compare within a row's source.
Local LLM (llama.cpp / Ollama)
| Setting | Result | Source |
|---|---|---|
| Llama 2 7B F16, Generation |
37.1tok/s | llama.cpp community benchmarks (GitHub), Oct 19, 2026 M5 Max, 40-core GPU, 614GB/s (configure-to-order) llama-bench tg128 (generating 128 tokens), all layers on the GPU. |
| Llama 2 7B Q4_0, Generation |
120tok/s | llama.cpp community benchmarks (GitHub), Oct 19, 2026 M5 Max, 40-core GPU, 614GB/s (configure-to-order) llama-bench tg128 (generating 128 tokens), all layers on the GPU. No 32-core base-model figure yet. |
| Llama 2 7B Q4_0, Prompt processing |
3,220tok/s | llama.cpp community benchmarks (GitHub), Oct 19, 2026 M5 Max, 40-core GPU, 614GB/s (configure-to-order) llama-bench pp512 (reading a 512-token prompt), all layers on the GPU. |
| Llama 2 7B Q8_0, Generation |
72.4tok/s | llama.cpp community benchmarks (GitHub), Oct 19, 2026 M5 Max, 40-core GPU, 614GB/s (configure-to-order) llama-bench tg128 (generating 128 tokens), all layers on the GPU. |
Geekbench AI
| Setting | Result | Source |
|---|---|---|
| Core ML Neural Engine, Quantized |
64,480points | Macworld, Sep 21, 2026 Mac Studio M5 Max, CTO 18-core CPU, 40-core GPU, 128GB, 4TB Quantized score, from Macworld's Geekbench AI Neural Engine chart |
Blender Benchmark
| Setting | Result | Source |
|---|---|---|
| Standard run |
6,462samples/min | Blender Open Data, Oct 19, 2026 Apple M5 Max (GPU - 32 cores), Metal, Blender 5.2.0 Median score over the three scenes, 11 submitted runs |
| 40-core GPU |
8,156samples/min | Blender Open Data, Oct 19, 2026 Apple M5 Max (GPU - 40 cores), Metal, Blender 5.2.0 Median score over the three scenes, 88 submitted runs |
| CPU |
585samples/min | Blender Open Data, Oct 19, 2026 Apple M5 Max CPU, Blender 5.2.0 Median score over the three scenes, 76 submitted runs. Mac Studio and MacBook Pro together. |
Geekbench 6 CPU
| Setting | Result | Source |
|---|---|---|
| Multi-core |
30,384points | Macworld, Sep 21, 2026 Mac Studio M5 Max, CTO 18-core CPU, 40-core GPU, 128GB, 4TB From Macworld's Geekbench 6 chart |
| Single-core |
4,355points | Macworld, Sep 21, 2026 Mac Studio M5 Max, CTO 18-core CPU, 40-core GPU, 128GB, 4TB From Macworld's Geekbench 6 chart |
Geekbench 6 GPU
| Setting | Result | Source |
|---|---|---|
| Metal |
238,904points | Macworld, Sep 21, 2026 Mac Studio M5 Max, CTO 18-core CPU, 40-core GPU, 128GB, 4TB From Macworld's Geekbench 6 Compute chart |
Sources
- Blender Open Data (results database)
- Macworld (review)
- llama.cpp community benchmarks (GitHub) (results database)
As an Amazon Associate we earn from qualifying purchases, and we may earn from other shop links. Amazon's price is on Amazon; Best Buy's is read from Best Buy and shown with the time it was checked.