M1 Ultra 128GB vs M5 Max 48GB for local LLMs
11 reports on the M1 Ultra 128GB and 67 on the M5 Max 48GB
Which is faster for local LLMs?
On Qwen3 8B at 4-bit, the M1 Ultra 128GB has 81.2 generation tokens/s (one run); prompt processing 518.0 tokens/s (one run); the M5 Max 48GB has 104.2 generation tokens/s (2 runs); prompt processing 2947.6 tokens/s (2 runs). 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 33% higher on the M1 Ultra 128GB (819 GB/s on the M1 Ultra 128GB, 614 on the M5 Max 48GB). Across 3 shared models with generation measurements on both cards, the M1 Ultra 128GB's median is higher on 0 and the M5 Max 48GB's on 3.
- Memory
- 128GB unified
- Memory bandwidth
- 819 GB/s
- FP16 compute
- 21 TFLOPS
- Reports
- 11
- Memory
- 48GB unified
- Memory bandwidth
- 614 GB/s
- FP16 compute
- not on record
- Reports
- 67
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 | M1 Ultra 128GB | M5 Max 48GB |
|---|---|---|
| Qwen3 8B4-bit | 81.2one runPP 518.0 (one run)engine not statedcontext not stated | 104.22 runsPP 2947.6 (2 runs)llama.cpp, MLXcontext not stated |
| Qwen3 4B4-bit | 116.9one runPP 942.3 (one run)engine not statedcontext not stated | 172.92 runsPP 5239.2 (2 runs)llama.cpp, MLXcontext not stated |
| Gemma 4 8B4-bit | 69.7one runPP 1172.9 (one run)llama.cppcontext not stated | 110.5one runPP 4217.3 (one run)llama.cppcontext not stated |
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 | M1 Ultra 128GB | M5 Max 48GB |
|---|---|---|
| Qwen3 8B | 25.2 t/sFP16 | 30.5 t/sFP16 |
| Qwen3 4B | 50.4 t/sFP16 | 61.0 t/sFP16 |
| Gemma 4 8B | 25.2 t/sFP16 | 30.5 t/sFP16 |
| Qwen3.8 27B | 7.5 t/sFP16 | 22.3 t/sQ6_K |
| Qwen3.8 125B · 6B active | 34.0 t/sQ5_K_S | doesn't fit at these settings |
| Qwen3.6 35B · 3B active | 23.5 t/sFP16 | 88.4 t/sQ5_K_M |
| DeepSeek V4 Flash 284B · 13B active | 53.5 t/sQ2_K | doesn't fit at these settings |
| Qwen3.6 27B | 7.5 t/sFP16 | 22.3 t/sQ6_K |
| Gemma 4 26B · 4B active | 20.2 t/sFP16 | 43.2 t/sQ8_0 |
Change the context, RAM or model in the calculator: M1 Ultra 128GB · M5 Max 48GB
Every report on each card, with its source: M1 Ultra 128GB · M5 Max 48GB