M3 Ultra 512GB vs M5 Ultra 256GB for local LLMs
41 reports on the M3 Ultra 512GB and 30 on the M5 Ultra 256GB
Which is faster for local LLMs?
On Gemma 4 31B at 16-bit, the M3 Ultra 512GB has 9.2 generation tokens/s (one run); prompt processing 295.9 tokens/s (one run); the M5 Ultra 256GB has 13.9 generation tokens/s (one run); prompt processing 1237.0 tokens/s (one run). Those are community reports whose engines and settings differ, so the gap is a rough sign, not a controlled measure of the two cards. Memory bandwidth, which caps how fast a card can write, is 50% higher on the M5 Ultra 256GB (819 GB/s on the M3 Ultra 512GB, 1,229 on the M5 Ultra 256GB). Across 4 shared models with generation measurements on both cards, the M3 Ultra 512GB's median is higher on 1 and the M5 Ultra 256GB's on 3.
- Memory
- 512GB unified
- Memory bandwidth
- 819 GB/s
- FP16 compute
- 28.4 TFLOPS
- Reports
- 41
- Memory
- 256GB unified
- Memory bandwidth
- 1,229 GB/s
- FP16 compute
- not on record
- Reports
- 30
Measured on both
Generation and prompt processing speeds from plain runs (one device, one request, no speculative decoding, the whole model in the device's memory) of the same model size at the same bit class. Each side shows the engines and reported context behind its median. Context can mean a configured window or prompt depth, so these are community comparisons with differing settings. How to read these
| Model | M3 Ultra 512GB | M5 Ultra 256GB |
|---|---|---|
| Gemma 4 31B16-bit | 9.2one runPP 295.9 (one run)MLXcontext not stated | 13.9one runPP 1237.0 (one run)MLXcontext not stated |
| Qwen3.5 122B · 10B active4-bit | 49.5one runPP 784.5 (one run)MLXcontext not stated | 74.7one runPP 2710.5 (one run)MLXcontext not stated |
| Qwen3.8 125B · 6B active8-bit | 25.7one runllama.cpp32K context | 17.0one runPP 1069.0 (one run)llama.cpp250K context |
| Qwen3.8 27B4-bit | 31.8one runPP 391.2 (one run)MLXcontext not stated | 50.0one runPP 1800.0 (one run)oMLX8K context |
Estimated by the calculator
The quant the calculator recommends for each card and its estimated speed at a 32,768-token context with 32 GB of system RAM, from memory bandwidth and the model's shape. Where a model was also measured above, the measurement wins.
| Model | M3 Ultra 512GB | M5 Ultra 256GB |
|---|---|---|
| Gemma 4 31B | 6.9 t/sFP16 | 12.9 t/sFP16 |
| Qwen3.5 122B · 10B active | 8.3 t/sFP16 | 31.0 t/sQ8_0 |
| Qwen3.8 125B · 6B active | 13.6 t/sFP16 | 50.8 t/sQ8_0 |
| Qwen3.8 27B | 8.0 t/sFP16 | 14.9 t/sFP16 |
| Qwen3.6 35B · 3B active | 28.7 t/sFP16 | 53.5 t/sFP16 |
| DeepSeek V4 Flash 284B · 13B active | 22.6 t/sQ8_0 | 66.9 t/sQ4_K_M |
| Qwen3.6 27B | 8.0 t/sFP16 | 14.9 t/sFP16 |
| Gemma 4 26B · 4B active | 21.5 t/sFP16 | 40.1 t/sFP16 |
| DeepSeek V4.1 Flash 552B · 16B active | 30.0 t/sQ4_K_M | 89.4 t/sQ2_K |
| GLM-5.3 320B · 18B active | 7.4 t/sQ8_0 | 35.6 t/sQ3_K_M |
Change the context, RAM or model in the calculator: M3 Ultra 512GB · M5 Ultra 256GB
Every report on each card, with its source: M3 Ultra 512GB · M5 Ultra 256GB