Apple M3 Ultra
DesktopApple's workstation-class unified-memory SoC (Mac Studio, shipping 2025-03-12): up to 32 CPU cores and 80 GPU cores, unified memory starting at 96 GB and configurable to 512 GB, with 819 GB/s of memory bandwidth. Apple's own framing is that it can run LLMs with more than 600B parameters entirely in memory. Apple does not publish a TDP, so that column is left empty.
3
Run records
1
Models covered
1
Frameworks covered
L0 Self-reported × 3
Evidence mix
Typical measured performance
Published benchmark averages per model × framework × quantization on this hardware.
- llama.cpp · IQ2_XS Decode 70.2 · Prefill 526.5 ·
- llama.cpp · Q5_K_M Decode 31.9 · Prefill 239.3 ·
- llama.cpp · Q4_K_M Decode 39 · Prefill 292.5 ·
Fit check
Physics-based estimate from VRAM and memory bandwidth: pick a model to see which quantizations fit and the expected speed
Formula: tok/s = (bandwidth ÷ model size) × 0.55, adapted from llmfit; see run records for measured numbers
Key specs
- Vendor
- Apple
- Architecture
- M3 Ultra
- VRAM
- 512 GB
- Memory bandwidth
- 819 GB/s
- Release date
- 2025-03-12
Run records
3 configurations
| Model | Framework | Quantization | Decode | Gen time | Samples |
|---|---|---|---|---|---|
| Qwen3.8-27B | llama.cpp | IQ2_XS | 70.2 tok/s | — | 1 |
| Qwen3.8-27B | llama.cpp | Q5_K_M | 31.9 tok/s | — | 1 |
| Qwen3.8-27B | llama.cpp | Q4_K_M | 39 tok/s | — | 1 |