M3 Ultra 256GB vs M4 24GB for local LLMs
7 reports on the M3 Ultra 256GB and 3 on the M4 24GB
Which is faster for local LLMs?
On Qwen3.5 9B at 4-bit, the M3 Ultra 256GB runs at 107.5 tokens per second (one run) and the M4 24GB at 17.4 (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 6.8x higher on the M3 Ultra 256GB (819 GB/s on the M3 Ultra 256GB, 120 on the M4 24GB).
- Memory
- 256GB unified
- Memory bandwidth
- 819 GB/s
- FP16 compute
- 28.4 TFLOPS
- Reports
- 7
- Memory
- 24GB unified
- Memory bandwidth
- 120 GB/s
- FP16 compute
- 4.3 TFLOPS
- Reports
- 3
Measured on both
Median tokens per second 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 quant level. Engines and context lengths can still differ between the runs. How to read these
| Model | M3 Ultra 256GB | M4 24GB |
|---|---|---|
| Qwen3.5 9B4-bit | 107.5one run | 17.4one run |
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 256GB | M4 24GB |
|---|---|---|
| Qwen3.5 9B | 23.9 t/sFP16 | 10.1 t/sQ8_0 |
| Qwen3.8 27B | 8.0 t/sFP16 | 10.8 t/sQ2_K |
| Qwen3.8 125B · 6B active | 27.3 t/sQ8_0 | doesn't fit at these settings |
| DeepSeek V4 Flash 284B · 13B active | 36.0 t/sQ4_K_M | doesn't fit at these settings |
| Qwen3.6 35B · 3B active | 28.7 t/sFP16 | 35.8 t/sQ2_K |
| Qwen3.6 27B | 8.0 t/sFP16 | 10.8 t/sQ2_K |
| Gemma 4 26B · 4B active | 21.5 t/sFP16 | 24.1 t/sQ3_K_S |
Change the context, RAM or model in the calculator: M3 Ultra 256GB · M4 24GB
Every report on each card, with its source: M3 Ultra 256GB · M4 24GB