Apple · Apple M5 Ultra · 2026
Apple Mac Studio M5 Ultra 512GB for local AI
The largest memory pool in a single desktop, with about 1.2TB/s of bandwidth to feed it.
Updated · Speeds marked ~ are estimates; seemethodology
- Unified memory
- 512GB
- Bandwidth
- 1,228 GB/s
- Power
- 480W
- Runs well
- 15 models
M5 Ultra 512GB
Can the M5 Ultra 512GB run LLMs?
The M5 Ultra 512GB has 512GB of unified memory (about 384GB usable by the GPU by default) and 1,228 GB/s of memory bandwidth, which decides how big a model fits and how fast it generates text. It comfortably runs models up to about 320B parameters: the largest in our catalog is GLM-5.3-Flash at Q6_K (273GB with 8K context, ~25 tokens/sec).
Recommendations
Best models for the M5 Ultra 512GB
Best overall · also coding, reasoning, vision
GLM-5.3-FlashQ6_K · 273GB · ~25 tokens/sec
Runs wellRunner-up
DeepSeek V4 FlashFP4 · 168GB · ~45 tokens/sec
Runs wellFastest good model
gpt-oss-20bMXFP4 · 15GB · ~105 tokens/sec
Runs wellLargest model that fits
GLM-5.3Q3_K_M · 375GB · ~19 tokens/sec
Fits, tight
Compatibility
What runs on the M5 Ultra 512GB
Every current model in our catalog, at the best quantization that fits in the ~384GB the GPU can use with 8K tokens of context. "Partial offload" means part of the model runs from system RAM (we assume 64GB), which is much slower.
| Model | Parameters | Best quant | Memory | Speed | Verdict |
|---|---|---|---|---|---|
| Gemma 4 E4B | 8B | Q8_0 | 10 GB | ~64 tok/s | Runs well |
| Qwen3.5 9B | 9B | Q8_0 | 12 GB | ~57 tok/s | Runs well |
| Phi-4 | 14.7B | Q8_0 | 18 GB | ~41 tok/s | Runs well |
| gpt-oss-20b | 20.9B (3.6B active) | MXFP4 | 15 GB | ~105 tok/s | Runs well |
| Gemma 4 26B A4B | 26B (3.8B active) | Q8_0 | 30 GB | ~73 tok/s | Runs well |
| Qwen3.8 27B | 27B | Q8_0 | 32 GB | ~22 tok/s | Runs well |
| Gemma 4 31B | 31B | Q8_0 | 36 GB | ~20 tok/s | Runs well |
| Qwen3.6 35B-A3B | 35B (3B active) | Q8_0 | 41 GB | ~83 tok/s | Runs well |
| GLM-4.5-Air | 106B (12B active) | Q8_0 | 118 GB | ~28 tok/s | Runs well |
| Llama 4 Scout | 109B (17B active) | Q8_0 | 120 GB | ~21 tok/s | Runs well |
| gpt-oss-120b | 117B (5.1B active) | MXFP4 | 68 GB | ~93 tok/s | Runs well |
| Qwen3.5 122B-A10B | 122B (10B active) | Q8_0 | 134 GB | ~33 tok/s | Runs well |
| DeepSeek V4 Flash | 284B (13B active) | FP4 | 168 GB | ~45 tok/s | Runs well |
| GLM-5.3-Flash | 320B (18B active) | Q6_K | 273 GB | ~25 tok/s | Runs well |
| GLM-5.3 | 744B (40B active) | Q3_K_M | 375 GB | ~19 tok/s | Fits, tight |
Performance
M5 Ultra 512GB vs. alternatives on Qwen3.8 27B
Generation speed in tokens per second for a single user. Token generation is limited mostly by memory bandwidth, which is why bandwidth predicts these results so well.
RTX PRO 6000
Q8_0 · 96GB
~46tok/s
M5 Ultra 512GB
Q8_0 · 512GB
~22tok/s
M5 Ultra 256GB
Q8_0 · 256GB
~22tok/s
M3 Ultra 512GB
Q8_0 · 512GB
~15tok/s
Our take
Should you buy the M5 Ultra 512GB for AI?
The top Mac Studio pairs the M5 Ultra's 80-core GPU with 512GB of unified memory and Apple's quoted 1.2TB/s of bandwidth, about 50 percent more than the M3 Ultra it replaces. That is enough to run DeepSeek-class models with hundreds of billions of parameters at 4 bit on one quiet machine, after raising the default GPU memory limit. Apple says each M5 GPU core adds a Neural Accelerator, which speeds up prompt processing, the weak spot of earlier Macs. The 512GB configuration ships in late October 2026, and Apple had not published its price as of mid-October.
What we like
- 512GB of unified memory in one desktop
- About 1.2TB/s of memory bandwidth
- Quiet, efficient and simple to set up with MLX or llama.cpp
- Neural Accelerators improve prompt processing over earlier Macs
Watch out for
- Price not announced as of October 2026
- No CUDA, so some tools and fine-tuning workflows do not run
- Memory cannot be upgraded after purchase
Best for: Running the largest open MoE models locally · A quiet single-box alternative to multi-GPU servers
Specifications
M5 Ultra 512GB specs that matter for AI
- Memory
- 512GB LPDDR5X (unified)
- GPU-usable memory
- ~384GB by default
- Memory bandwidth
- 1,228 GB/s
- Architecture
- Apple M5 Ultra
- Power
- 480W (system)
- Interface
- Thunderbolt 5 (no PCIe slots)
- Released
- 2026
- Software
- Metal, MLX
macOS lets the GPU use about 75 percent of memory by default (about 384GB), and the iogpu.wired_limit_mb setting can raise it. Apple quotes 1.2TB/s for M5 Ultra.
Sources: Apple Mac Studio (M5 Max or M5 Ultra) tech specs, MacRumors: Mac Studio with M5 Ultra and 512GB launching in October, Apple Mac Studio power consumption and thermal output, ModelPiper: raising the Metal memory ceiling with iogpu.wired_limit_mb
Cross-shop
Alternatives to the M5 Ultra 512GB
Frequently asked questions
How much VRAM does the M5 Ultra 512GB have?
The Apple Mac Studio M5 Ultra 512GB has 512GB of unified memory (about 384GB usable by the GPU by default) with 1,228 GB/s of memory bandwidth.
What is the largest LLM the M5 Ultra 512GB can run?
Fully in GPU memory, the largest model in our catalog that runs comfortably is GLM-5.3-Flash (320B (18B active)) at Q6_K, using about 273GB with 8K context at ~25 tokens/sec.
What is the best LLM to run on the M5 Ultra 512GB?
Our top pick is GLM-5.3-Flash at Q6_K, which fits in 512GB with room for 8K of context and runs at ~25 tokens/sec. See our full list of the best models for the M5 Ultra 512GB for coding, reasoning, and vision picks.
Is the M5 Ultra 512GB good for local AI in 2026?
The largest memory pool in a single desktop, with about 1.2TB/s of bandwidth to feed it.