M4 Pro LLM benchmark
No signed M4 Pro submissions yet. Check the configuration guidance below, or contribute a reproducible measurement from your own machine.
Specs vendor specs, not measured
- Unified memory (max configuration)
- 64 GB
- Memory bandwidth
- 273 GB/s
- Launch date
- 2024-10-30
Sources: support.apple.com, support.apple.com, apple.com. Vendor-published figures, accessed 2026-10-14.
M4 Pro: start with your installed memory
Apple lists the 2024 Mac mini with M4 Pro at 24 GB, configurable to 48 GB or 64 GB, and with a 16- or 20-core GPU. The 64 GB figure is a configuration ceiling, not the memory in every M4 Pro. For a MacBook Pro, check its own configuration. Apple Mac mini specifications
Model weights share memory with macOS, other applications, the runtime and the context cache. A parameter count alone cannot tell you whether a model fits. Start the calculator with your installed capacity, then choose the model, quantization and context length:
The calculator is a rough screen, including a memory reserve. It does not establish runtime compatibility, model quality or measured speed. For a model near the limit, verify the actual downloaded file and test the prompt length and concurrent applications you intend to use.
A published M4 Pro speed reference
The llama.cpp Apple Silicon summary lists these Llama 2 7B Q4_0 results with build 8e672ef. They are historical, third-party reference figures, checked September 16, 2026. They are not signed llm-speed submissions or tests of current models.
| GPU cores | Prompt tok/s | Generation tok/s |
|---|---|---|
| 16 | 364.06 | 49.64 |
| 20 | 439.78 | 50.74 |
These two rows show a larger difference in prompt processing than generation. They do not measure time to first token, long-context performance or the effect of upgrading memory. Do not extrapolate them to Gemma 4, Qwen3.8 or MLX. Compare your actual workload before paying for a higher specification.
For longer documents or coding-agent prompts, read how prompt processing differs from generation.
Contribute a signed M4 Pro benchmark
Run on YOUR hardware to populate this page: pipx install llm-speed && llm-speed bench
$ pipx install llm-speed && llm-speed bench
Community folklore on M4 Pro
14 unverified claims extracted from Reddit/HN comments. Lower trust than signed runs above; every row links to the source.
- communityconfidence 70%
70.00tok/s — Qwen3-Coder on M4 Pro via mlx
our signed data: M4 Pro · Qwen3-Coder
“urpose-built for code, MoE architecture so only 3B params active per token. Fits in 36GB with room for 16-32K context. On M4 Pro MLX I get ~70 tok/s with it. If you also want a general-purpose model to keep alongside it, **Qwen3.5-35B-A3B** is the same MoE architecture, similar …”
- communityconfidence 60%
7.24tok/s — qwen2.5 on M4 Pro via ollama
our signed data: M4 Pro · qwen2.5
“2.5 right up until you hit the 72B parameter size. Due to VRAM it falls apart. The M4 max and M4 Pro could run the 72B, but at 8.8t/s and 7.24t/s. This is certainly better than running on a CPU (which is what happens with the 4090), but it's still too slow to be worth a $4k pu…”
- communityconfidence 60%
8.80tok/s — qwen2.5 on M4 Pro via ollama
our signed data: M4 Pro · qwen2.5
“lly at qwen2.5 right up until you hit the 72B parameter size. Due to VRAM it falls apart. The M4 max and M4 Pro could run the 72B, but at 8.8t/s and 7.24t/s. This is certainly better than running on a CPU (which is what happens with the 4090), but it's still too slow to be wor…”
- communityconfidence 60%
72.00tok/s — GPT OSS 20B on M4 Pro via mlx
our signed data: M4 Pro · GPT OSS 20B
“edge frozen as of 2024-06, handled rescaling and converting a recipe into metric that Qwen 3 30B 3AB 2507 completely hosed, churns at about 72 tok/s on a MacBook M4 Pro Max 4 bit MLX. And it ALMOST got a side scrolling shooter working, whereas the 120B didn't, and Qwen 3 didn't …”
- communityconfidence 60%
14.50tok/s — Gemma 3 27b on M4 Pro via mlx
our signed data: M4 Pro · Gemma 3 27b
“Mac Mini M4 Pro (20c GPU) 64GB unified RAM; Gemma 3 27b with MLX 14.5t/s power usage almost 70W (including connected keyboard and mouse). So more efficient then even the Ryzen 395 AI (if those results are accurat”
- communityconfidence 60%
55.00tok/s — Qwen3 30b a3b on M4 Pro via mlx
our signed data: M4 Pro · Qwen3 30b a3b
“Binned M4 Pro/48GB owner here since November--current daily driver is Qwen3 30b a3b 8-bit MLX @ 55t/s ymmv, but I like it a lot and it flies.”
- communityconfidence 60%
11.37tok/s — Qwen3 30b-a3b on M4 Pro via mlx
our signed data: M4 Pro · Qwen3 30b-a3b
“That's about three times as fast as my binned M4 Pro/48GB: with Qwen3 30b-a3b 8-bit MLX, I got 180 seconds to first token and 11.37t/s with the same size lorem ipsum prompt. That tracks really well with the 3X memory bandwidth difference.”
- communityconfidence 55%
60.50tok/s — on M4 Pro via lm-studio FP16
our signed data: M4 Pro
“32 GB DDR5|CUDA 12 llama.cpp (LM Studio)|59.1 tok/s|0.02 s|\-| |M4 Pro|16 GPU cores, MacBook Pro 14”, 48 GB unified memory|MLX (LM Studio)|60.5 tok/s 👑|0.31 s|3.69| # Super Interesting Notes: **1. The neural accelerators didn't make much of a difference. Here's why!** * First …”
- communityconfidence 55%
10.00tok/s — on M4 Pro via mlx bf16
our signed data: M4 Pro
“end half a day downloading each one, hoping someone with more hands-on experience can point me in the right direction. Update: Tried 6Bit- 10 tok/sec. Not that great. https://preview.redd.it/xms2wfztttwg1.png?width=3022&format=png&auto=webp&s=71579f674ec522e75673e4746b8adaa9…”
- communityconfidence 55%
45.00tok/s — Qwen3-30B-A3B on M4 Pro Q6_K
our signed data: M4 Pro · Qwen3-30B-A3B
“Qwen3-30B-A3B is very fast for how capable it is. I’m getting about 45 t/s on my unbinned M4 Pro Mac Mini with 64GB Ram. In my experience, it’s good all around, but not as good as GLM4-32B 0414 Q6_K on one-shoting”
Request a run on M4 Pro
No measured run yet for these pairs. Vote for the one you want; the most-requested runs go first. See all requests.
We never sell or share your email. Privacy