Apple · Apple M3 Ultra · 2025
Apple Mac Studio M3 Ultra 512GB for local AI
The original 512GB Mac for giant models, now available only used and at volatile prices.
Updated · Speeds marked ~ are estimates; seemethodology
- Unified memory
- 512GB
- Bandwidth
- 819 GB/s
- Power
- 480W
- Runs well
- 15 models
M3 Ultra 512GB
Typically $9,499 at launch
Can the M3 Ultra 512GB run LLMs?
The M3 Ultra 512GB has 512GB of unified memory (about 384GB usable by the GPU by default) and 819 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, ~17 tokens/sec).
Recommendations
Best models for the M3 Ultra 512GB
Best overall · also coding, reasoning, vision
GLM-5.3-FlashQ6_K · 273GB · ~17 tokens/sec
Runs wellRunner-up
DeepSeek V4 FlashFP4 · 168GB · ~32 tokens/sec
Runs wellFastest good model
gpt-oss-20bMXFP4 · 15GB · ~80 tokens/sec
Runs wellLargest model that fits
GLM-5.3Q3_K_M · 375GB · ~13 tokens/sec
Fits, tight
Compatibility
What runs on the M3 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 | ~44 tok/s | Runs well |
| Qwen3.5 9B | 9B | Q8_0 | 12 GB | ~39 tok/s | Runs well |
| Phi-4 | 14.7B | Q8_0 | 18 GB | ~28 tok/s | Runs well |
| gpt-oss-20b | 20.9B (3.6B active) | MXFP4 | 15 GB | ~80 tok/s | Runs well |
| Gemma 4 26B A4B | 26B (3.8B active) | Q8_0 | 30 GB | ~54 tok/s | Runs well |
| Qwen3.8 27B | 27B | Q8_0 | 32 GB | ~15 tok/s | Runs well |
| Gemma 4 31B | 31B | Q8_0 | 36 GB | ~14 tok/s | Runs well |
| Qwen3.6 35B-A3B | 35B (3B active) | Q8_0 | 41 GB | ~62 tok/s | Runs well |
| GLM-4.5-Air | 106B (12B active) | Q8_0 | 118 GB | ~20 tok/s | Runs well |
| Llama 4 Scout | 109B (17B active) | Q8_0 | 120 GB | ~14 tok/s | Runs well |
| gpt-oss-120b | 117B (5.1B active) | MXFP4 | 68 GB | ~70 tok/s | Runs well |
| Qwen3.5 122B-A10B | 122B (10B active) | Q8_0 | 134 GB | ~23 tok/s | Runs well |
| DeepSeek V4 Flash | 284B (13B active) | FP4 | 168 GB | ~32 tok/s | Runs well |
| GLM-5.3-Flash | 320B (18B active) | Q6_K | 273 GB | ~17 tok/s | Runs well |
| GLM-5.3 | 744B (40B active) | Q3_K_M | 375 GB | ~13 tok/s | Fits, tight |
Performance
M3 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.
M5 Ultra 512GB
Q8_0 · 512GB
~22tok/s
M5 Ultra 256GB
Q8_0 · 256GB
~22tok/s
M3 Ultra 512GB
Q8_0 · 512GB
~15tok/s
Measured
M3 Ultra 512GB benchmark results
| Model | Quant | Runtime | Generation | Prompt | Source |
|---|---|---|---|---|---|
| Llama 2 7B Chat | Q4_0 | llama.cpp (Metal) | 92 tok/s | 1470 tok/s | llama.cpp discussion #4167: Performance of llama.cpp on Apple Silicon M-series · October 2026 |
| Llama 2 7B Chat | Q8_0 | llama.cpp (Metal) | 64 tok/s | 1490 tok/s | llama.cpp discussion #4167: Performance of llama.cpp on Apple Silicon M-series · October 2026 |
| Llama 2 7B Chat | F16 | llama.cpp (Metal) | 40 tok/s | 1540 tok/s | llama.cpp discussion #4167: Performance of llama.cpp on Apple Silicon M-series · October 2026 |
Our take
Should you buy the M3 Ultra 512GB for AI?
The 2025 Mac Studio with M3 Ultra and 512GB was the first desktop that could hold DeepSeek-class models at 4 bit, with 819 GB/s of memory bandwidth. It made the Mac Studio a common choice for people running the largest open models without a GPU server. Apple removed the 512GB option in March 2026 during the memory shortage and replaced the line with M5 Ultra models in September. Prompt processing is slower than on the M5 Ultra, which adds GPU Neural Accelerators and about 50 percent more bandwidth.
What we like
- 512GB of unified memory
- 819 GB/s bandwidth
- Quiet and efficient for its capacity
Watch out for
- Discontinued, with volatile used prices
- Slow prompt processing on long inputs
- No CUDA support
Best for: Running DeepSeek-class models at 4 bit · Buyers who find one used below M5 Ultra pricing
Specifications
M3 Ultra 512GB specs that matter for AI
- Memory
- 512GB LPDDR5 (unified)
- GPU-usable memory
- ~384GB by default
- Memory bandwidth
- 819 GB/s
- Architecture
- Apple M3 Ultra
- Power
- 480W (system)
- Interface
- Thunderbolt 5 (no PCIe slots)
- Released
- 2025
- Launch price
- $9,499
- 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.
Sources: Apple Mac Studio (2025) tech specs, MacRumors: Apple cuts more Mac Studio and Mac mini RAM options, Apple Mac Studio power consumption and thermal output
Cross-shop
Alternatives to the M3 Ultra 512GB
Frequently asked questions
How much VRAM does the M3 Ultra 512GB have?
The Apple Mac Studio M3 Ultra 512GB has 512GB of unified memory (about 384GB usable by the GPU by default) with 819 GB/s of memory bandwidth.
What is the largest LLM the M3 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 ~17 tokens/sec.
What is the best LLM to run on the M3 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 ~17 tokens/sec. See our full list of the best models for the M3 Ultra 512GB for coding, reasoning, and vision picks.
Is the M3 Ultra 512GB good for local AI in 2026?
The original 512GB Mac for giant models, now available only used and at volatile prices.