Apple's Mac Studio gets M5 Max and M5 Ultra chips with 1.2TB/s memory bandwidth

Mac Studio refreshed with M5 Max and M5 Ultra

Apple has launched a new version of the Mac Studio desktop built around its M5 Max and M5 Ultra processors, with the top configuration delivering 1.2TB/s of memory bandwidth. The details come from Apple's Mac Studio product page, which reached the Hacker News front page on August 26.

Two chips, one workstation

According to Apple, the M5 Max is aimed at creators, developers and engineers who need heavy CPU, GPU and on-device AI compute. The company claims up to 2.5x faster CPU performance, 4.9x faster GPU performance and 10.7x faster AI performance compared with the M1 Max. The chip includes a third-generation ray-tracing engine, supports up to five displays, and can play up to 19 simultaneous streams of 8K ProRes 422 video at 30 fps.

The M5 Ultra is built by fusing two M5 Max chips into a single system-on-a-chip. Apple says it offers the highest CPU core count of any Mac, the most powerful GPU the company has put in Apple silicon, and its most capable chip for AI work. Compared with the M1 Ultra, Apple claims up to 2.4x faster CPU, 4.7x faster GPU and 9.8x faster AI performance. It drives up to eight displays and handles up to 33 simultaneous 8K ProRes 422 streams at 30 fps.

One detail worth flagging: all of Apple's performance comparisons on the page are against the M1 generation, not the more recent M4 chips. No M4-to-M5 figures appear in the material surfaced.

Pitched as a local AI machine

The marketing leans heavily on running AI workloads on device. Apple says Mac Studio can run large language models locally through LM Studio Bionic, deploy open-weight models including Gemma, Kimi and Mistral, generate images with Stable Diffusion and FLUX, and build agentic workflows that execute entirely on the machine. A tool called exo can cluster several AI workstations together for larger jobs.

Apple also cites workload-specific gains: faster time-to-first-token in LM Studio, faster machine-learning training for Foundry Nuke's CopyCat feature, faster Magic Mask in Blackmagic DaVinci Resolve Studio, faster DNA basecalling in Oxford Nanopore's MinKNOW, quicker Redshift renders, faster Xcode builds, and higher ray-traced frame rates in Cyberpunk 2077: Ultimate Edition.

The company frames privacy as a core part of the pitch, emphasising on-device processing and its Private Cloud Compute architecture, which it says lets Apple Intelligence draw on larger server-side models without making user data accessible to Apple or others.

Leasing and software extras

The page also introduces Apple Upgrade, a leasing programme that spreads the cost of a new Mac over 24 or 36 monthly payments with an upgrade option at the end of the term.

On the software side, Apple highlights a reworked Siri AI with a dedicated app and Visual Intelligence, natural-language Shortcuts automation, and Safari features such as automatic tab grouping by topic and notifications when a watched page changes.

Pricing, memory configurations and availability dates were not included in the material surfaced, so those details remain unconfirmed.

Why it matters

Memory bandwidth is the binding constraint for local LLM inference, since token generation speed depends largely on how fast weights can be streamed from memory. A 1.2TB/s figure, paired with the large unified memory typical of Apple's pro desktops, is what makes running big models locally viable without discrete GPUs or per-token cloud costs. The Ultra's dual-die design also shows how Apple plans to keep scaling performance past the limits of a single chip. For developers and studios weighing local hardware against rented cloud GPU capacity, this release is Apple's most direct argument yet that serious AI work can happen on the desk rather than in the data centre.