AI Hardware

Local AI and GPU hardware2026-10-28· By Franz Senn

MacBook Pro with PROLED XDR: touch, tandem OLED and 128 GB of memory in a laptop

On 28 October 2026 Apple introduced the MacBook Pro with PROLED XDR Display. Pre-orders opened the same day, and it ships from Friday 6 November. It is the first MacBook Pro with touch and with tandem OLED, and the memory tiers go up to 128 GB. For local language models, memory bandwidth decides the speed and capacity only decides whether a model fits. We have no measurements of our own for the machine, and we know of no independent ones yet.

The MacBook Pro with PROLED XDR Display is the laptop Apple introduced with the M5 Pro and M5 Max chips. It has a tandem OLED with two layers of self-emitting pixels, up to 2,000 nits peak brightness, a touch sensor built into the display panel and, depending on the build, up to 128 GB of unified memory. In Germany it costs from 3,299 euros including VAT, and shipping starts on 6 November.

The press release is dated 28 October 2026, and prices and dates in this post stand on that day. We have no measurements of our own for this machine. (Quelle: Wikimedia Commons, trademark: Apple)

Prices according to Apple Germany

Four configurations carry a price in the release (Apple Germany press release). Every figure is a manufacturer claim, and we have not checked them against anything.

Model	Chip	Price from, euros incl. VAT	Education price from, euros
14-inch	M5 Pro	3,299	3,129
14-inch	M5 Max	5,119	4,859
16-inch	M5 Pro	3,899	3,669
16-inch	M5 Max	5,599	5,249

Table 1: entry prices by model and chip, as of the release on 28 October 2026, source Apple Germany.

The 128 GB tier does not appear in this price list. On the order page, the presets we retrieved on 5 November show configurations up to 48 GB (order page). What the configurator asks for 128 GB we have not checked.

Further details from release and data sheet: three Thunderbolt 5 ports, a 12 MP camera with Center Stage, storage up to 16 TB, battery life of up to 24 hours according to Apple, and a weight of 1.38 kilograms (14-inch) and 1.82 kilograms (16-inch) with M5 Pro. With M5 Max the data sheet gives 1.40 and 1.84 kilograms.

Memory and bandwidth in the data sheet

The technical specifications list the bandwidth, the press release does not.

Chip	GPU cores	Memory, GB	Bandwidth, GB/s
M5 Pro	16	24 or 48	307
M5 Pro	20	24, 48 or 64	307
M5 Max	32	36	460
M5 Max	40	48, 64 or 128	614

Table 2: memory tiers and bandwidth, manufacturer figures from the Apple data sheet, retrieved 5 November 2026. Not re-measured by us.

So 128 GB exists only with the M5 Max with 40 GPU cores and 614 GB/s.

PROLED XDR: what is new about the display

Apple describes a tandem OLED with two layers of self-emitting pixels, plus algorithms in the display engine and a factory calibration of every panel. The figures from release and data sheet: up to 2,000 nits peak brightness, up to 1,000 nits sustained at full screen, 10,000,000:1 contrast and ProMotion from 10 to 120 Hz. Apple says the nano-texture glass is now etched directly into the cover glass.

For the predecessor the data sheet gives 1,600 nits peak for HDR content and up to 1,000 nits for SDR outdoors. What is new, then, is mainly that the 2,000 nits peak also applies to SDR, there outdoors according to the data sheet.

Four manufacturer claims, no measurement. For the new model Apple gives 1,000 nits as sustained full-screen brightness. (Quelle: Apple, technical specifications MacBook Pro with PROLED XDR Display (DE) and MacBook Pro 14-inch M4 (DE), retrieved 5 November 2026)

The second new thing is touch. The sensor, Apple says, sits in the display panel and not as a layer on top, and the Dynamic Island arrives, on a Mac for the first time. Touch adds to keyboard and trackpad and does not replace them. How it feels in daily work is in no document.

128 GB says nothing about speed

When generating text, the model has to be read through memory once per token. That is why bandwidth limits the speed, as we measured with four RTX 5090s and their DRAM bandwidth. Capacity decides whether a model fits together with its context.

By arithmetic a model with 70 billion parameters at 4 bit takes about 35 GB, and a model with 200 billion already takes about 100 GB. The key-value cache comes on top and grows with context length. In its press release Apple writes that with the M5 Max, language models with hundreds of billions of parameters can run entirely on the device. The release names no quantisation, context length or speed for that, and we cannot check the statement.

(Quelle: Our own arithmetic from parameter count times bits per weight, bandwidths from the Apple data sheet)

Some orientation comes from the September reviews of the Mac Studio with M5 Ultra. We summarised them in our post on the first measured numbers. The figures below are not ours. They come from Alex Grankin's roundup (video, 22 September 2026) and apply, per the video, to Qwen3.8-27B at 4 bit.

Chip	Bandwidth per Apple, GB/s	Generation speed, tokens/s
M4 Max	up to 546	23
M5 Max	614	30
M3 Ultra	819	30
M5 Ultra	1,200	46 to 54

Table 3: third-party figures. Qwen3.8-27B, 4 bit, roundup by Alex Grankin of 22 September 2026, machines and software not broken down. Bandwidths are manufacturer figures.

Alex Ziskind (video) measured the M5 Ultra's bandwidth with a GPU microbenchmark: 1,039 of 1,200 GB/s. Whether the M5 Max values come from a laptop or a desktop is not clear from the roundup, and a laptop's cooling can change the result over sustained load.

Laptop or desktop

Anyone who wants 128 GB of memory for language models has two build types to choose from. The laptop is lighter and mobile. As a desktop, the Mac Studio has more room for cooling. By our price table of 6 September 2026 the Mac Studio with M5 Max and 128 GB cost 5,859 euros, and the M5 Ultra tier with 256 GB cost 12,429 euros. Those are street prices as of that day. We do not know the price of the laptop with 128 GB yet.

Our post on the reviews has the M5 Ultra reading the prompt about twice as fast as the M5 Max and writing at roughly 1.5 times. At equal memory, mobility is the only argument for the laptop.

Our own production runs on four RTX 5090s (how the AI stack is built), and we work out what such an operation costs per task in AI cost per task. If you want to run a model in-house, our offer is described under Corporate LLM stack: AI on your own hardware, and more measurement series are collected in the hub Local AI and GPU hardware.

What stays open until the first measurement

Four points need the machine in hand:

The measured memory bandwidth of M5 Pro and M5 Max.
Tokens per second with a fixed reference model and fixed parameters.
Behaviour after 30 minutes of sustained load, because a laptop throttles differently from a desktop.
The price of the 128 GB tier and a delivery date from a retailer.

Until then the machine is a promise in a data sheet for local language models, and what the documents support is mainly the display. Anyone who needs a laptop with a lot of memory for models should wait for the first independent measurements after 6 November before confirming the order.