Apple’s M5 Max shipped in early 2026 with 460 GB/s memory bandwidth and up to 128GB unified memory. Reviews called it “the best machine for local LLMs you can buy.” They’re not wrong about the bandwidth. But bandwidth isn’t the whole story.

I run local AI inference on an AMD Ryzen AI Max+ 395 (Strix Halo architecture, 128GB LPDDR5x-8000 unified memory, ~256 GB/s bandwidth). Let me tell you what the benchmarks don’t say.

The Numbers

M5 Max AMD Strix Halo
Memory bandwidth 460 GB/s ~256 GB/s
Max unified memory 128 GB 128 GB
GPU architecture 32-core Apple GPU RDNA 3.5 (40 CUs)
GPU memory usable up to 128 GB up to ~115 GB (GTT)
Software stack Metal + MLX ROCm + Vulkan + llama.cpp
Price (fully loaded) ~$7,000+ ~$900 (Beelink GTR9 Pro)

On bandwidth-bound inference, the M5 Max wins. For a 7B Q4_K_M model (roughly 4.4GB), it will push noticeably more tokens per second. Benchmarks I’ve seen put it around 80–100 tok/s for Llama 3.1 8B. My machine does around 55–65 tok/s on the same class of model.

For medium models — 30–40B range — the M5 Max’s bandwidth advantage shrinks proportionally as the model size grows relative to bandwidth. Both machines slow down. The M5 Max still leads.

Where It Gets Interesting: Big Models

Here’s the thing about 460 GB/s bandwidth with 36GB of RAM: you top out at roughly 20–22GB usable for a model (leaving headroom for OS, context, KV cache). That means the M5 Max 36GB variant maxes out at about a 20B Q4_K_M model running comfortably. For 30B+ models you start compressing aggressively or hitting page boundaries.

The M5 Max 128GB variant solves this — but at $6,000–7,000+ retail.

My machine (the Beelink GTR9 Pro at ~$900) also has 128GB unified memory, with ~115GB accessible to the GPU via GTT (Graphics Translation Table). I run Qwen3.5-35B-A3B at around 22 tok/s. I run our 122B MoE model at around 8–10 tok/s. The M5 Max 36GB can’t load either of these without severe quantization or layer offloading.

For the same ~$900 price bracket, there’s simply nothing on Apple’s side. The cheapest M5 Max is in a $2,499+ MacBook.

Bandwidth vs. Capacity: A Real Tradeoff

Memory bandwidth determines how fast you push tokens through a model that fits in memory. Memory capacity determines which models you can run at all.

The M5 Max is a sprinter. Excellent at medium-sized models, screaming fast on small ones, limited in the heavyweight division.

Strix Halo is more of a long-distance runner. Slower peak speed, but it can carry models the M5 Max 36GB can’t even lift. And at $900 vs $3,000+, you can afford the trade-off.

The Ecosystem Question

This is where it gets uncomfortable if you’re in the Apple ecosystem.

The M5 Max uses Metal and MLX. MLX is excellent software — Apple’s team has done great work. But it’s proprietary infrastructure, macOS-only, and Apple decides what gets optimized and when.

My AMD setup runs llama.cpp with Vulkan. It runs on Linux. It runs the same code that runs in data centers. When llama.cpp adds Llama 4 support, or Mamba, or whatever comes next, I have it the same day. No waiting for MLX to catch up, no platform gate, no macOS required.

I’ve had llama.cpp support for new architectures within hours of a major model release. That matters more than you’d think.

Practical Verdict

If you’re buying a laptop and want the best local LLM experience money can buy, the M5 Max 128GB variant ($6,000+) is genuinely impressive. Fast, silent, excellent battery, polished software.

If you want to run the biggest open-weight models with a fixed budget under $1,500, AMD Strix Halo is the better choice right now. The GTR9 Pro at ~$900 is almost offensively good value.

And if you want to run both 7B quick-response models and 120B+ quality models on the same machine — there’s currently only one architecture that supports that without a second device: Strix Halo’s 128GB unified with full GPU access.

The M5 Max is faster per token. My machine runs more of them.

I run llama.cpp with Vulkan backend on the GTR9 Pro. Benchmarks for our specific models (Qwen3.5-35B, qwen3-coder, devstral) on this hardware are in other posts. Bandwidth figures: M5 Max 460 GB/s per Apple spec; Strix Halo ~256 GB/s per AMD LPDDR5x-8000 spec and community benchmarks.