MacBook Pro with M5 Pro and M5 Max: Apple targets people running LLMs locally

On March 3, Apple announced the new MacBook Pro with the M5 Pro and M5 Max chips. In past years I would have read an announcement like this thinking about video rendering and compile times. This time, Apple’s own text makes it clear which audience it wants to win: people who run language models locally. For anyone who writes software, that is the part that really matters.

What changed in the chips

According to Apple, the M5 Pro and M5 Max ship with up to an 18-core CPU, split into 6 “super cores” and 12 performance cores. Apple calls the super core “the world’s fastest CPU core” and promises up to 30% faster CPU performance than the M4 Pro and M4 Max. On the GPU side, it claims up to 50% higher graphics performance than the previous generation.

The most relevant change, though, is architectural: every GPU core now includes a Neural Accelerator. That is what backs the AI numbers Apple highlighted. The company says LLM prompt processing is up to 3.9x faster on the M5 Pro compared with the M4 Pro, and up to 4x faster on the M5 Max compared with the M4 Max. Against the M1 Pro and M1 Max, the claimed gains reach 6.9x and 6.7x respectively. For AI image generation, Apple cites up to 3.7x (M5 Pro) and 3.8x (M5 Max) over the M4 generation.

Memory and bandwidth: what matters for local models

Anyone who has tried to run a large model on a laptop knows the bottleneck is rarely raw compute alone. It is memory. The model has to fit, and token generation speed depends heavily on how fast the weights can be read.

Here are the numbers:

M5 Pro: up to 64GB of unified memory, with up to 307GB/s of bandwidth.
M5 Max: up to 128GB of unified memory, with up to 614GB/s of bandwidth.

With 128GB shared between CPU and GPU, you can load models that simply do not fit in the VRAM of a consumer graphics card. That mix of capacity and bandwidth is what makes the Mac attractive for local inference, and the M5 Max doubles the bandwidth of the M5 Pro.

Note that the metric Apple highlighted is prompt processing, meaning the phase where the model reads the input context. In coding agent workflows, which dump entire files into the context, that phase carries a lot of weight. Token generation speed depends more on memory bandwidth, and the announcement does not give a specific gain figure for it.

The rest of the machine

Apple says the SSD is up to 2x faster, reaching 14.5GB/s read and write. Base storage went up: 1TB on the M5 Pro and 2TB on the M5 Max. Battery life is rated at up to 24 hours, with fast charging up to 50% in 30 minutes using a 96W or higher USB-C adapter.

For connectivity there are three Thunderbolt 5 ports, HDMI with support for up to 8K, an SDXC card slot and MagSafe 3. The new N1 wireless chip brings Wi-Fi 7 and Bluetooth 6. The M5 Pro drives up to two external displays, and the M5 Max up to four.

Pricing and availability

In the United States, prices start at:

14-inch with M5 Pro: $2,199
16-inch with M5 Pro: $2,699
14-inch with M5 Max: $3,599
16-inch with M5 Max: $3,899

The 14-inch model with the base M5 also now starts with 1TB, at $1,699. Pre-orders opened on March 4 in 33 countries and regions, and the machines reach stores starting March 11.

My take

For developers, the practical question is: is it worth paying for this to run models locally? The answer depends on the size of the model you want to use. The 64GB on the M5 Pro already handles quantized mid-sized models well. If the goal is to experiment with bigger models or keep several loaded at once, only the M5 Max with 128GB will do, and that is where the price climbs fast.

The numbers also call for caution. Every performance gain cited here comes from Apple, measured under conditions Apple chose. I will wait for independent tests, with real models and real tools, before trusting the “up to 4x”.

Still, the direction is clear. Apple is no longer treating local AI as a side effect of the GPU. Putting neural accelerators in every GPU core and talking openly about LLM prompt processing shows that, for Apple, the developer laptop has also become an inference workstation. For anyone who wants privacy, offline work or simply a way out of the monthly API bill, that is good news.