Apple · Local AI machine
Apple Mac Studio (M5 Ultra, 256 GB)
Apple machine: Apple M5 Ultra (30-core CPU, 64-core GPU), 256 GB memory, 192 GB for the GPU, 1200 GB/s memory speed, about 18.4 tok/s on a 70B model (estimate).
86.1Score
- #2 of 41 ranked
- #2 of 10 Apple machines
Why this score
The score adds up these parts. Each part has a weight. A part counts as the average of its sources, pulled toward 80 when there are few sources.
AI power
50% weightcounts as 90.9Average of 2 sources: 96.3
Experts
30% weightcounts as 82.3Average of 4 sources: 82.9
Users
20% weightcounts as 80.0No sources found yet. No data yet. This part counts as 80.
Good: 4 reviews and buyer ratings.
80 is the field average. A part counts as 80 + (average − 80) × N ÷ (N + 1), where N is its number of sources. One source counts half of the gap from 80. No sources count as 80.
Price
€10999 at Apple Store Österreich · Austrian shop
Apple AT list price, lowest 256 GB configuration: M5 Ultra 30-core CPU / 64-core GPU, 256 GB, 1 TB SSD. The 36-core CPU / 80-core GPU chip with 256 GB, 1 TB costs EUR 12,429. Geizhals Austria lowest offer for the 64-core config is EUR 10,800.00 (Cyberport.at, computeruniverse.at, pre-order).
Prices checked 8 Oct 2026.
Specs
- Chip
- Apple M5 Ultra (30-core CPU, 64-core GPU)
- CPU
- 30 cores (10 Super + 20 Performance)
- Memory
- 256 GB
- GPU memory
- 192 GBmacOS default Metal GPU memory cap of about 75% of unified memory (community-documented in devnote-override-macos-metal-vram-cap and mlx-lm issue 883; Apple publishes no share). 75% x 256 GB = 192 GB. Raising the cap with sysctl iogpu.wired_limit_mb works but is unsupported by Apple, so it is not applied. Bandwidth 1.2 TB/s (Apple spec page).
- Memory speed
- 1200 GB/s
- FP16 TFLOPS
- —
- NPU
- —
- OS
- macOS
- Software
- MLX
- Ethernet
- 10 Gbit/s
- Thunderbolt / USB4
- TB5
- QSFP port
- No
- Idle power
- 9 W
- Load power
- 385 W
- Noise
- —
- Volume
- 3.7 L
- Weight
- 3.6 kg
- SSD
- 1 TB
- Release
- 2026-09
Measured
- Qwen3.5-122B-A10B, not stated (MLX)
- 80 tok/s writing · LM Studio (MLX), M5 Ultra 256 GB, 36-core CPU / 80-core GPU
- Qwen3.8 27B (dense), 4-bit MLX
- 48 tok/s writing, 1701 tok/s prompt · oMLX 0.7.0.dev2 (MLX), M5 Ultra 256 GB, 6,091-token prompt
- Flash-Next (MacStories' name; size not stated), oQ4e (oMLX quant)
- 90.7 tok/s writing, 2733 tok/s prompt · oMLX 0.7.0.dev2 (MLX), M5 Ultra 256 GB, 4K context
Notes
Fastest memory (1.2 TB/s) and the only small box here that holds 256 GB, so reviewers saw about 80 tok/s on a 122B-A10B MoE model and 1,700-2,700 tok/s prompt reading; Apple documents a Neural Accelerator in every GPU core. Very expensive (EUR 6.6k-11k+), memory is soldered, and the 512 GB option is not orderable yet.
Apple's spec page lists the 256 GB option next to the 36-core CPU / 80-core GPU chip, but Apple's AT buy page also sells 256 GB with the 64-core GPU (EUR 10,999); the price uses the cheaper one. Apple's power page lists the 36c/80c/512 GB/16 TB configuration (idle 9 W, max 385 W), not the base config. Tom's Hardware and PCMag UK scores were read via the Geizhals review table; the links go to the reviews. Prices are Apple AT list prices: the lower Geizhals offers in each price note were not linked in the research, and Geizhals could not be read again.
More reading:macstories.netbgr.competapixel.commacdailynews.comtechradar.comchannelpartner.deithinkdiff.com