Apple M5 Ultra
台式机Apple 工作站级统一内存 SoC,2026-08-25 随新款 Mac Studio 发布:最高 36 核 CPU / 80 核 GPU,最高 512GB 统一内存,内存带宽 1.2 TB/s(约 1200 GB/s,比上代高 50%)。是目前单机可装载最大模型的消费级平台之一。Apple 不公布 TDP,该列留空。
69
运行记录
25
覆盖模型
2
覆盖框架
L0 自报 × 7 · L2 跨框架验证 × 62
证据分布
典型性能表现
同硬件上各模型 × 框架 × 量化的已上架实测均值。
- llama.cpp · IQ3_S Decode 153.37 · Prefill 580.58 ·
- Ollama · IQ2_M Decode 156.39 · Prefill 558.43 ·
- llama.cpp · Q5_K_M Decode 84.63 · Prefill 432.9 ·
- Ollama · Q3_K_S Decode 125.07 · Prefill 471.87 ·
- llama.cpp · IQ2_M Decode 381.7 · Prefill 1687.3 ·
- Ollama · IQ2_XS Decode 337.8 · Prefill 1438.9 ·
- llama.cpp · Q4_K_M Decode 130.96 · Prefill 547.83 ·
- Ollama · Q4_K_M Decode 106.35 · Prefill 353.45 ·
- llama.cpp · IQ1_S Decode 324.79 · Prefill 1453.99 ·
- llama.cpp · Q5_K_S Decode 181.42 · Prefill 831.96 ·
- Ollama · Q2_K Decode 230.73 · Prefill 1077.9 ·
- llama.cpp · Q8_0 Decode 76.11 · Prefill 389.07 ·
- Ollama · Q4_K_S Decode 118.22 · Prefill 450.95 ·
- llama.cpp · Q3_K_M Decode 90.54 · Prefill 497.72 ·
- llama.cpp · Q5_K_S Decode 71.27 · Prefill 325.67 ·
- Ollama · IQ4_XS Decode 81.44 · Prefill 327.07 ·
- llama.cpp · Q3_K_S Decode 117.38 · Prefill 386.02 ·
- llama.cpp · Q5_K_S Decode 83.32 · Prefill 309.36 ·
- Ollama · Q6_K Decode 62.32 · Prefill 341.91 ·
- llama.cpp · IQ2_XS Decode 89.58 · Prefill 565.47 ·
- Ollama · Q2_K Decode 74.38 · Prefill 283.32 ·
- Ollama · Q3_K_S Decode 60.73 · Prefill 250.32 ·
- llama.cpp · IQ3_XS Decode 91.11 · Prefill 454.84 ·
- llama.cpp · Q4_K_S Decode 81.06 · Prefill 387.21 ·
- Ollama · IQ3_S Decode 85.88 · Prefill 388.53 ·
- Ollama · Q6_K Decode 53.29 · Prefill 218.93 ·
- llama.cpp · Q5_K_S Decode 169.24 · Prefill 662.43 ·
- Ollama · Q3_K_M Decode 206.74 · Prefill 725.54 ·
- llama.cpp · IQ1_S Decode 190.95 · Prefill 992.52 ·
- Ollama · Q3_K_L Decode 144 · Prefill 660.08 ·
- llama.cpp · IQ1_S Decode 131.5 · Prefill 444.06 ·
- Ollama · Q3_K_S Decode 84 · Prefill 452.07 ·
- llama.cpp · Q8_0 Decode 165.05 · Prefill 859.2 ·
- Ollama · IQ2_M Decode 328.08 · Prefill 1800.71 ·
- Ollama · IQ3_XS Decode 260.88 · Prefill 1003.67 ·
- llama.cpp · IQ2_M Decode 130.42 · Prefill 681.87 ·
- Ollama · Q8_0 Decode 57.41 · Prefill 209.64 ·
- llama.cpp · Q4_K_M Decode 102.09 · Prefill 450.54 ·
- llama.cpp · Q6_K Decode 73.25 · Prefill 373.92 ·
- Ollama · IQ3_XS Decode 108 · Prefill 446.71 ·
- Ollama · Q4_K_S Decode 102.76 · Prefill 333.29 ·
- llama.cpp · Q3_K_L Decode 65.89 · Prefill 258.39 ·
- llama.cpp · Q8_0 Decode 37.99 · Prefill 165.65 ·
- Ollama · Q2_K Decode 75.92 · Prefill 280.65 ·
- Ollama · Q8_0 Decode 38.85 · Prefill 140.28 ·
- llama.cpp · IQ3_XS Decode 105.53 · Prefill 434.57 ·
- Ollama · Q3_K_L Decode 86.92 · Prefill 466.52 ·
- Ollama · Q5_K_S Decode 66.58 · Prefill 239.33 ·
- llama.cpp · Q3_K_S Decode 134.78 · Prefill 545.26 ·
- llama.cpp · Q5_K_S Decode 104.75 · Prefill 395.8 ·
Fit 判定
按显存与内存带宽做物理估算:选择模型,查看各量化能否装下及预计速度
估算公式:tok/s =(带宽 ÷ 模型占用)× 0.55,借鉴 llmfit;实测以运行记录为准
核心规格
- 厂商
- Apple
- 架构
- M5 Ultra
- 显存
- 512 GB
- 显存带宽
- 1200 GB/s
- 发布日期
- 2026-08-25
运行记录
共 50 组配置
| 模型 | 框架 | 量化 | Decode | 生成耗时 | 样本数 |
|---|---|---|---|---|---|
| gemma-4-E4B-it | llama.cpp | IQ3_S | 153.4 tok/s | — | 1 |
| gemma-4-E4B-it | Ollama | IQ2_M | 156.4 tok/s | — | 1 |
| Ministral-3-8B-Instruct-2512 | llama.cpp | Q5_K_M | 84.6 tok/s | — | 1 |
| Ministral-3-8B-Instruct-2512 | Ollama | Q3_K_S | 125.1 tok/s | — | 1 |
| iFlytek Spark-X2.5-4B | llama.cpp | IQ2_M | 381.7 tok/s | — | 1 |
| iFlytek Spark-X2.5-4B | Ollama | IQ2_XS | 337.8 tok/s | — | 1 |
| NVIDIA-Nemotron-3.5-Bolt-E4B-4GB | llama.cpp | Q4_K_M | 131 tok/s | — | 1 |
| NVIDIA-Nemotron-3.5-Bolt-E4B-4GB | Ollama | Q4_K_M | 106.3 tok/s | — | 1 |
| Flint-40B-A2.5B | llama.cpp | IQ1_S | 324.8 tok/s | — | 1 |
| Flint-40B-A2.5B | llama.cpp | Q5_K_S | 181.4 tok/s | — | 1 |
| Flint-40B-A2.5B | Ollama | Q2_K | 230.7 tok/s | — | 1 |
| Raon-1-mini-preview | llama.cpp | Q8_0 | 76.1 tok/s | — | 1 |
| Raon-1-mini-preview | Ollama | Q4_K_S | 118.2 tok/s | — | 1 |
| mammoth-1 | llama.cpp | Q3_K_M | 90.5 tok/s | — | 1 |
| mammoth-1 | llama.cpp | Q5_K_S | 71.3 tok/s | — | 1 |
| mammoth-1 | Ollama | IQ4_XS | 81.4 tok/s | — | 1 |
| elephant-1 | llama.cpp | Q3_K_S | 117.4 tok/s | — | 1 |
| elephant-1 | llama.cpp | Q5_K_S | 83.3 tok/s | — | 1 |
| elephant-1 | Ollama | Q6_K | 62.3 tok/s | — | 1 |
| Qwen3.8-27B | llama.cpp | IQ2_XS | 89.6 tok/s | — | 2 |
| Qwen3.8-27B | Ollama | Q2_K | 74.4 tok/s | — | 1 |
| Qwen3.8-27B | Ollama | Q3_K_S | 60.7 tok/s | — | 1 |
| Qwen3.6-27B | llama.cpp | IQ3_XS | 91.1 tok/s | — | 1 |
| Qwen3.6-27B | llama.cpp | Q4_K_S | 81.1 tok/s | — | 1 |
| Qwen3.6-27B | Ollama | IQ3_S | 85.9 tok/s | — | 1 |
| Qwen3.6-27B | Ollama | Q6_K | 53.3 tok/s | — | 1 |
| Qwen3.6-35B-A3B | llama.cpp | Q5_K_S | 169.2 tok/s | — | 1 |
| Qwen3.6-35B-A3B | Ollama | Q3_K_M | 206.7 tok/s | — | 1 |
| Qwen3.5-4B | llama.cpp | IQ1_S | 190.9 tok/s | — | 1 |
| Qwen3.5-4B | Ollama | Q3_K_L | 144 tok/s | — | 1 |
| Qwen3.5-27B | llama.cpp | IQ1_S | 131.5 tok/s | — | 1 |
| Qwen3.5-27B | Ollama | Q3_K_S | 84 tok/s | — | 1 |
| Qwen3.5-0.8B | llama.cpp | Q8_0 | 165.1 tok/s | — | 1 |
| Qwen3.5-0.8B | Ollama | IQ2_M | 328.1 tok/s | — | 1 |
| Qwen3.5-0.8B | Ollama | IQ3_XS | 260.9 tok/s | — | 1 |
| Qwen3.8-9B | llama.cpp | IQ2_M | 130.4 tok/s | — | 1 |
| Qwen3.8-9B | Ollama | Q8_0 | 57.4 tok/s | — | 1 |
| Qwen3-8B | llama.cpp | Q4_K_M | 102.1 tok/s | — | 1 |
| Qwen3-8B | llama.cpp | Q6_K | 73.3 tok/s | — | 1 |
| Qwen3-8B | Ollama | IQ3_XS | 108 tok/s | — | 1 |
| Qwen3-8B | Ollama | Q4_K_S | 102.8 tok/s | — | 1 |
| gemma-4-31B-it | llama.cpp | Q3_K_L | 65.9 tok/s | — | 1 |
| gemma-4-31B-it | llama.cpp | Q8_0 | 38 tok/s | — | 1 |
| gemma-4-31B-it | Ollama | Q2_K | 75.9 tok/s | — | 1 |
| gemma-4-31B-it | Ollama | Q8_0 | 38.9 tok/s | — | 1 |
| gemma-4-12B-it | llama.cpp | IQ3_XS | 105.5 tok/s | — | 1 |
| gemma-4-12B-it | Ollama | Q3_K_L | 86.9 tok/s | — | 1 |
| gemma-4-12B-it | Ollama | Q5_K_S | 66.6 tok/s | — | 1 |
| Llama-3.1-8B-Instruct | llama.cpp | Q3_K_S | 134.8 tok/s | — | 1 |
| Llama-3.1-8B-Instruct | llama.cpp | Q5_K_S | 104.8 tok/s | — | 1 |