Apple's August 25 newsroom post prices a Mac Studio with M5 Max from $2,499 and one with M5 Ultra from $5,499. The post says the Max goes up to 128GB of unified memory and the Ultra up to 512GB. It does not print a sticker for the 128GB Max. These closer figures are US list readings of the configurator. They are not a sentence in that post.

The spec sheet is where the two machines separate. 128GB exists only on the M5 Max with an 18-core CPU, a 40-core GPU, and 614GB/s of memory bandwidth. The base Max, a 32-core GPU with 36GB, is 460GB/s, and it cannot be ordered with 128GB. The Ultra at $5,499 is the base chip: a 30-core CPU with 10 super cores and 20 performance cores, a 64-core GPU, 96GB of memory, a 1TB SSD, and 1.2TB/s of bandwidth.

John Gruber published the configurator ladder the same day, in hundred-dollar steps. The 40-core Max is $3,100, and that price already includes 48GB. From there, 64GB is $400 and 128GB is $2,000. Storage on the Max starts at 512GB, and a 1TB SSD is $300. Those steps put a 128GB Max with 512GB of storage at $5,100, and the same machine with 1TB at $5,400. localaiprices reads the 512GB-storage version at $5,099. The page says that figure was checked by hand on apple.com, last confirmed on October 16, and that the listing still showed $5,099 on October 29. The dollar under Gruber's figure is the same rounding as Apple's $2,499 base against his $2,500.

Matched at 1TB, the pair is about $100 apart: about $5,400 against the $5,499 Ultra. A September 7 video from Adam Talks Tech describes that pair the same way, about $5,400 and about $100 under the Ultra. The 128GB Max with only the base 512GB SSD is the $5,099 machine, about $400 under the Ultra, and it is not the storage-matched pair.

The Max reaches this price because 128GB is locked to the larger GPU. On August 26, 9to5Mac wrote that the configurator shows the 40-core GPU as a $300 option, then changes the memory as well, another $300, so the step is $600. 9to5Mac called that memory change a move from 32GB to 48GB. The spec sheet lists the base Max at 36GB, not 32GB. Gruber's two Max prices, $2,500 with 36GB and $3,100 with 48GB, are $600 apart.

Bandwidth does not climb when the memory does. The spec sheet's larger Ultra, a 36-core CPU and an 80-core GPU, is still 1.2TB/s. VideoCardz's October 29 table puts that chip at $6,799 with 96GB and a 1TB SSD. More CPU cores, more GPU cores, the same memory bandwidth, and $1,300 above the base Ultra. It is not the machine in this comparison. On the chips that are, the Max has 6 super cores and 12 performance cores. The Neural Engine is 16 cores on the Max and 32 on the Ultra. Apple does not print a TOPS number for either chip.

The media engines differ, and that is a spec-sheet fact rather than a model result. The Max has one video decode engine, two encode engines, and two ProRes engines. The Ultra has two, four, and four. Apple's newsroom says that is twice the encode and decode blocks, and that an M5 Ultra can play up to 33 streams of 8K ProRes 422 at 30 fps. The front panel differs too. An M5 Max has two USB-C ports at up to 10Gb/s. An M5 Ultra has two Thunderbolt 5 ports. Both have an SDXC slot. The rear is shared: four Thunderbolt 5 ports, 10Gb Ethernet, two USB-A ports, HDMI 2.1, and a headphone jack. Display support is five on the Max and eight on the Ultra. Weight is 2.7 kg on the Max and 3.6 kg on the Ultra.

What fits is a count, not a run on either machine. A dense model at 4-bit is half a gigabyte per billion parameters, before context. The token ceiling is the memory bandwidth divided by that weight. The ceilings below take Apple's 1.2TB/s as 1,200GB/s, which is just under twice 614GB/s. A heavier packaged file lowers the ceiling. Memory left over is what the system and the context share, so a real run sits under the ceiling.

Qwen3 14B is 7GB at this count. The ceiling is about 88 tokens a second on the Max and about 171 on the Ultra. Qwen3 32B is 16GB, about 38 tokens a second on the Max and 75 on the Ultra. A dense 70-billion-parameter model is 35GB, about 18 and about 34. A packaged Q4_K_M file runs heavier than the count, about 9 to 11GB for Qwen3 14B and about 20GB for Qwen3 32B. On a 20GB file the ceiling falls to about 31 tokens a second on the Max and 60 on the Ultra.

All three of those files fit in 96GB, and they fit in 128GB. For one Qwen3 14B, one Qwen3 32B, or one 70B model, the Ultra is the faster machine at about the same price. Two 70B files are 70GB at the count. They still fit the 96GB machine. Three are 105GB. They do not fit in 96GB. They fit in 128GB, with little room left beside them.

That leftover band is where the Max is the one to buy. A dense 4-bit file over 96GB, and still short of 128GB, leaves the Ultra unable to hold it and still leaves the Max room for the system and the context. At this count that is a model just over 192 billion parameters and short of 256 billion. Few dense models sit there. A file of 96GB of weights would fill the Ultra and leave nothing beside it. A mixture-of-experts model is a different test. It fits or it does not by the size of the file you download. The active-parameter count in the name does not decide it.

A 405-billion-parameter dense model at this count is about 203GB. It fits neither studio. Apple sells 256GB only on the 36-core, 80-core Ultra. VideoCardz lists that machine, with 1TB of storage, at $10,799. The 512GB configuration is another price again, on that same large chip. Neither figure is the $100 decision.

For a model that already fits, the extra 32GB does not buy tokens. The bandwidth does. The 128GB Max is the buy when the weights, or the files kept in memory together, land between 96GB and 128GB, and the jump to a 256GB Ultra is more machine than the work needs.

Buy the 96GB Ultra for one 14B, 32B, or 70B model. Buy the 128GB Max only for a file between those two sizes. Buy neither for a 405B model at this 4-bit count.