# M5 Ultra

> OnPrem World · https://onpremworld.com/en/hardware/apple-m5-ultra · Last updated: 2026-10-14 · CC BY 4.0

1.2 TB/s and 96–512 GB of unified memory in the Mac Studio: 256 GB and up holds models like DeepSeek-V4-Flash that no single GPU here can. Apple says the 512 GB option ships in late October 2026.

## Specifications

|  |  |
|---|---|
| Memory | 96–512 GB LPDDR5X |
| Memory bandwidth | 1200 GB/s |
| Launch | 2026-09 |
| Launch price | US$5,499 |
| Architecture | Apple M5 Ultra (quad-die: two dual-die M5 Max joined by UltraFusion; 3 nm) |
| Software | Metal / MLX |

## What it can run (512 GB, 8K)

Estimates, not benchmarks. See the methodology.

| Model | Parameters | Best quant | Total | tok/s (est.) |
|---|---|---|---|---|
| DeepSeek-V4-Flash | 284B | Q8_0 | 295.6 GiB | 56 |
| Qwen3.8-Flash-Next | 180B | FP16 | 352.7 GiB | 65 |
| Qwen3.5-122B-A10B | 122B | FP16 | 239.3 GiB | 39 |
| Mistral Small 4 119B (2603) | 119B | FP16 | 233.4 GiB | 60 |
| gpt-oss-120b | 117B | FP16 | 229.6 GiB | 76 |
| Llama 4 Scout 17B-16E Instruct | 109B | FP16 | 215.2 GiB | 23 |
| GLM-4.5-Air | 106B | FP16 | 209.2 GiB | 33 |
| Qwen3-Coder-Next (80B-A3B) | 80B | FP16 | 157.1 GiB | 130 |
| Llama 3.3 70B Instruct | 70B | FP16 | 139.9 GiB | 5.6 |
| Qwen3.5-35B-A3B | 35B | FP16 | 69.1 GiB | 130 |
| Qwen3.6-35B-A3B | 35B | FP16 | 69.1 GiB | 130 |
| Qwen3-32B | 32.8B | FP16 | 66.6 GiB | 12 |
| DeepSeek-R1-Distill-Qwen-32B | 32.76B | FP16 | 66.6 GiB | 12 |
| Gemma 4 31B IT | 30.7B | FP16 | 61.6 GiB | 13 |
| Qwen3-30B-A3B (2507) | 30.5B | FP16 | 60.9 GiB | 118 |
| Qwen3-Coder-30B-A3B-Instruct | 30.5B | FP16 | 60.9 GiB | 118 |
| GLM-4.7-Flash (30B-A3B) | 30B | FP16 | 59.6 GiB | 130 |
| Qwen3.5-27B | 27B | FP16 | 53.8 GiB | 14 |
| Qwen3.6-27B | 27B | FP16 | 53.8 GiB | 14 |
| Qwen3.8-27B | 27B | FP16 | 53.8 GiB | 14 |
| Gemma 3 27B IT | 27B | FP16 | 54.3 GiB | 14 |
| Gemma 4 26B A4B IT | 25.2B | FP16 | 50.1 GiB | 103 |
| Mistral Small 3.2 24B Instruct (2506) | 24B | FP16 | 48.7 GiB | 16 |
| Magistral Small 1.2 (2509) | 24B | FP16 | 48.7 GiB | 16 |
| Devstral Small 2 24B Instruct (2512) | 24B | FP16 | 48.7 GiB | 16 |
| gpt-oss-20b | 21B | FP16 | 41.8 GiB | 108 |
| Phi-4 (14B) | 14B | FP16 | 29.4 GiB | 28 |
| Gemma 3 12B IT | 12B | FP16 | 24.8 GiB | 33 |
| Gemma 4 12B Unified IT | 11.95B | FP16 | 24.2 GiB | 33 |
| Qwen3.5-9B | 9B | FP16 | 18.4 GiB | 43 |
| Qwen3-8B | 8.2B | FP16 | 17.7 GiB | 48 |
| DeepSeek-R1-0528-Qwen3-8B | 8.19B | FP16 | 17.6 GiB | 48 |
| Llama 3.1 8B Instruct | 8B | FP16 | 17.1 GiB | 49 |
| Gemma 4 E4B IT | 8B | FP16 | 16.4 GiB | 49 |
| Llama 3.2 3B Instruct | 3.21B | FP16 | 7.7 GiB | 121 |
| Kimi K2.6 | 1000B | Doesn't fit | – | – |
| DeepSeek-R1-0528 (671B) | 671B | Doesn't fit | – | – |

## Prices

Latest price at each shop, as listed on the day we checked.

| Region | Shop | Listing | Price | Per GB | Checked |
|---|---|---|---|---|---|
| Taiwan | Apple | Mac Studio，M5 Ultra 晶片，30 核心 CPU、64 核心 GPU，96GB 統一記憶體，1TB 儲存裝置 (Price shown in Apple Store configurator for this configuration) | [NT$199,900](https://www.apple.com/tw/shop/buy-mac/mac-studio) | NT$2,082/GB | 2026-10-14 |
| US | Apple | Mac Studio, M5 Ultra chip, 30-core CPU, 64-core GPU, 96GB memory, 1TB storage (Price shown in Apple Store configurator for this configuration) | [US$5,499](https://www.apple.com/shop/buy-mac/mac-studio) | US$57/GB | 2026-10-14 |

Taiwan value: about 160.8 tok/s on Llama 3.1 8B Instruct (Q4_K_M, 8K), or 0.8 tok/s per NT$1,000.

US value: about 160.8 tok/s on Llama 3.1 8B Instruct (Q4_K_M, 8K), or 2.9 tok/s per US$100.

All prices: https://onpremworld.com/en/prices.md

## Frequently asked questions

### Which LLMs can the M5 Ultra run?

With 512 GB and an 8K context, the largest models that fit include: DeepSeek-V4-Flash, Qwen3.8-Flash-Next, Qwen3.5-122B-A10B, Mistral Small 4 119B (2603), gpt-oss-120b. See the full table below.

### How fast is the M5 Ultra for local LLMs?

Generation speed is mostly set by memory bandwidth (1200 GB/s). For Llama 3.1 8B Instruct at FP16, expect roughly 49 tokens/s; bigger models are slower.

### Can the M5 Ultra run 70B models?

Yes. Llama 3.3 70B Instruct fits on the M5 Ultra (512 GB) at FP16 with an 8K context, using about 139.9 GiB, at an estimated 5.6 tokens/s.

## Sources

- https://www.apple.com/mac-studio/specs/
- https://en.wikipedia.org/wiki/Apple_M5
- https://www.apple.com/newsroom/2026/09/the-new-mac-mini-and-mac-studio-are-available-today/
- https://www.apple.com/newsroom/2026/08/apple-introduces-new-mac-studio-with-m5-max-and-m5-ultra/
- https://www.apple.com/newsroom/2026/08/apple-introduces-m6-and-m5-ultra-for-a-big-leap-in-performance-and-ai-compute/