GPUs & Macs › Apple M2 Ultra
Apple M2 Ultra: local LLM speed
Llama 2 7B (Q4_0) generates 36.1 tokens/s on the Apple M2 Ultra (Vulkan), and processes prompts at 199 tokens/s.
Measured results
| Model | Generate tok/s | Prompt tok/s | Quant | Backend | Test | Source |
|---|---|---|---|---|---|---|
| Llama 2 7B | 36.1 | 199 | Q4_0 | Vulkan FA on | tg128 | llama.cpp #10879 details |
| Llama 2 7B | 34.3 | 206 | Q4_0 | Vulkan FA off | tg128 | llama.cpp #10879 details |
| Gemma 3 4B IT (QAT) | 134 | 2,924 | Q4_0 | Metal FA on | tg32 | llama.cpp benches details |
| gpt-oss 20B | 130 | 2,713 | MXFP4 | Metal FA on | tg32 | llama.cpp benches details |
| gpt-oss 120B | 85.6 | 1,649 | MXFP4 | Metal FA on | tg32 | llama.cpp benches details |
| Qwen2.5 Coder 7B | 79.7 | 1,566 | Q8_0 | Metal FA on | tg32 | llama.cpp benches details |
| Qwen3 Coder 30B-A3B Instruct | 79.0 | 2,453 | Q8_0 | Metal FA on | tg32 | llama.cpp benches details |
| GLM-4.7-Flash | 59.6 | 1,629 | Q8_0 | Metal FA on | tg32 | llama.cpp benches details |
What fits on the Apple M2 Ultra
Estimates for this chip are scaled from a measured run: Qwen2.5 Coder 7B Q8_0 at 79.7 tok/s (Metal, source). Fit assumes the whole model in GPU memory with ~8K context.
With 16 GB unified memory
| Model | Best quant that fits | File | Expected tok/s |
|---|---|---|---|
| Qwen3.8-27B | UD-IQ2_XXS | 7.3 GB | 88.8 estimate |
| gemma-4-12B-it | Q6_K | 9.8 GB | 65.9 estimate |
| Qwen3.5-9B | Q8_0 | 9.5 GB | 67.7 estimate |
| Qwen3.5-4B | Q8_0 | 4.5 GB | 144 estimate |
| Qwen3.6-35B-A3B | UD-IQ1_M | 10.0 GB | 375 estimate |
| gemma-4-26B-A4B-it | UD-IQ2_M | 10.0 GB | 209 estimate |
| MiMo-V2.6-Distill-Qwen-9B | Q8_0 | 9.5 GB | 67.6 estimate |
| gemma-4-E4B-it | Q8_0 | 8.2 GB | 78.8 estimate |
With 36 GB unified memory
| Model | Best quant that fits | File | Expected tok/s |
|---|---|---|---|
| Qwen3.8-27B | UD-Q6_K | 22.0 GB | 29.4 estimate |
| gemma-4-12B-it | Q8_0 | 12.7 GB | 50.9 estimate |
| Qwen3.5-9B | Q8_0 | 9.5 GB | 67.7 estimate |
| Qwen3.5-4B | Q8_0 | 4.5 GB | 144 estimate |
| Qwen3.6-35B-A3B | UD-Q5_K_S | 24.9 GB | 151 estimate |
| gemma-4-26B-A4B-it | UD-Q6_K | 23.2 GB | 90.5 estimate |
| MiMo-V2.6-Distill-Qwen-9B | Q8_0 | 9.5 GB | 67.6 estimate |
| gemma-4-E4B-it | Q8_0 | 8.2 GB | 78.8 estimate |
With 64 GB unified memory
| Model | Best quant that fits | File | Expected tok/s |
|---|---|---|---|
| Qwen3.8-27B | Q8_0 | 29.0 GB | 22.2 estimate |
| gemma-4-12B-it | Q8_0 | 12.7 GB | 50.9 estimate |
| Qwen3.5-9B | Q8_0 | 9.5 GB | 67.7 estimate |
| Qwen3.5-4B | Q8_0 | 4.5 GB | 144 estimate |
| Qwen3.6-35B-A3B | Q8_0 | 36.9 GB | 102 estimate |
| gemma-4-26B-A4B-it | Q8_0 | 26.9 GB | 78.1 estimate |
| MiMo-V2.6-Distill-Qwen-9B | Q8_0 | 9.5 GB | 67.6 estimate |
| gemma-4-E4B-it | Q8_0 | 8.2 GB | 78.8 estimate |
With 128 GB unified memory
| Model | Best quant that fits | File | Expected tok/s |
|---|---|---|---|
| Qwen3.8-27B | Q8_0 | 29.0 GB | 22.2 estimate |
| gemma-4-12B-it | Q8_0 | 12.7 GB | 50.9 estimate |
| Qwen3.5-9B | Q8_0 | 9.5 GB | 67.7 estimate |
| Qwen3.5-4B | Q8_0 | 4.5 GB | 144 estimate |
| Qwen3.6-35B-A3B | Q8_0 | 36.9 GB | 102 estimate |
| Qwen3.8-Flash-Next | UD-IQ3_XXS | 82.0 GB | 7.9 estimate |
| gemma-4-26B-A4B-it | Q8_0 | 26.9 GB | 78.1 estimate |
| MiMo-V2.6-Distill-Qwen-9B | Q8_0 | 9.5 GB | 67.6 estimate |
Similar hardware
Apple M1 (7-core GPU) · Apple M1 (8-core GPU) · Apple M1 Pro (14-core GPU) · Apple M1 Pro (16-core GPU) · Apple M1 Max (24-core GPU) · Apple M1 Max (32-core GPU) · Apple M1 Ultra (48-core GPU) · Apple M1 Ultra (64-core GPU)