趋近智
Apple M4 Pro (48GB)
Apple M4 Pro (48GB) 采用统一内存架构。为了在避免系统内存压力的情况下确保本地大模型稳定推理，本评测以 75% 可用配额（36 GB）及 273 GB/s 统一内存带宽进行建模。
总显存
48 GB
可用显存上限
36 GB （75% 可用预算）
显存带宽
273 GB/s
最大推理模型级别
32B 模型 (Q4)
热设计功耗
85W
显存占用随上下文长度变化曲线
随上下文长度增长的显存需求评估。当曲线超过参考线时将导致显存不足（OOM）。
量化精度:
Apple M4 Pro (48GB) 最佳运行大模型与速度矩阵
主流开源大模型在 Apple M4 Pro (48GB) 上的推理速度、显存兼容性与生成吞吐表现。
排序:
上下文长度:
KV 缓存:
| 排名 | 模型 | 参数量 | 量化精度 | 预估 TPS | TTFT | 可运行 |
|---|---|---|---|---|---|---|
#33 | Qwen 3.8 27B | 27B | INT4 Q8 | 14 tok/s 8 tok/s | ~29.4s ~32.0s | |
#48 | Agnes 3.0 Flash | 33B | INT4 Q8 | 12 tok/s 6 tok/s | ~35.8s ~42.2s | |
#72 | K2 Horizon MoVA 36B A4B | 36B (4B active) | INT4 Q8 | 61 tok/s 38 tok/s | ~5.8s ~6.8s | |
#83 | Qwen3.5-27B | 27B | INT4 Q8 | 14 tok/s 8 tok/s | ~29.4s ~32.0s | |
#92 | Sarvam-30B | 32B (2.4B active) | INT4 Q8 | 81 tok/s 57 tok/s | ~4.0s ~4.4s | |
#94 | Qwen3.6 35B A3B | 35B (3B active) | INT4 Q8 | 72 tok/s 46 tok/s | ~4.7s ~5.5s | |
#96 | Muse Glimmer 30B | 30B | INT4 Q8 | 13 tok/s 7 tok/s | ~32.6s ~35.5s | |
#100 | Phi-4 Reasoning Plus | 14B | INT4 Q8 FP16 | 19 tok/s 13 tok/s 7 tok/s | ~15.5s ~15.5s ~16.9s | |
#101 | MiMo V2 Flash | 15B (309B active) | INT4 Q8 FP16 | 1 tok/s 1 tok/s 1 tok/s | ~321.8s ~322.4s ~380.6s | |
#108 | Gemma 4 26B A4B | 25.2B (3.8B active) | INT4 Q8 | 66 tok/s 45 tok/s | ~5.3s ~5.3s | |
#109 | Gemma 4 12B | 11.95B | INT4 Q8 FP16 | 22 tok/s 15 tok/s 8 tok/s | ~13.2s ~13.3s ~13.3s | |
#113 | Qwen3.5-9B | 9B | INT4 Q8 FP16 | 39 tok/s 23 tok/s 12 tok/s | ~10.0s ~10.0s ~10.1s | |
#114 | Qwen3.5-35B-A3B | 35B (3B active) | INT4 Q8 | 72 tok/s 46 tok/s | ~4.7s ~5.5s | |
#118 | K2 Horizon 7B | 7B | INT4 Q8 FP16 | 48 tok/s 29 tok/s 15 tok/s | ~7.8s ~7.8s ~7.9s | |
#128 | NVIDIA Nemotron 3 Nano 30B-A3B | 3.5B (30B active) | INT4 Q8 FP16 | 12 tok/s 7 tok/s 4 tok/s | ~31.6s ~31.6s ~31.8s | |
#129 | Gemma 4 31B | 30.7B | INT4 Q8 | 13 tok/s 7 tok/s | ~33.3s ~36.3s | |
#131 | Nemotron 3.5 Lightning | 30B (3B active) | INT4 Q8 | 27 tok/s 22 tok/s | ~4.6s ~5.4s | |
#132 | Qwen3.5-4B | 4B | INT4 Q8 FP16 | 73 tok/s 47 tok/s 26 tok/s | ~4.6s ~4.6s ~4.6s | |
#133 | Qwen3-30B-A3B | 30B (3B active) | INT4 Q8 | 73 tok/s 49 tok/s | ~4.6s ~5.0s | |
#136 | GPT-OSS 20B | 21B (3.6B active) | INT4 Q8 | 30 tok/s 26 tok/s | ~4.9s ~4.9s | |
#146 | Granite 4.2 8B | 8B | INT4 Q8 FP16 | 43 tok/s 26 tok/s 14 tok/s | ~8.9s ~8.9s ~9.0s | |
#150 | Gemma 4 E4B | 8B | INT4 Q8 FP16 | 43 tok/s 26 tok/s 14 tok/s | ~8.9s ~8.9s ~9.0s | |
#151 | Nemotron 3.5 Lightning 30B A3B | 30B (3B active) | INT4 Q8 | 73 tok/s 49 tok/s | ~4.6s ~5.0s | |
#153 | Qwen3-32B | 32B | INT4 Q8 | 12 tok/s 7 tok/s | ~34.7s ~37.8s | |
#155 | Mistral Large 3 | 41B (675B active) | INT4 | 1 tok/s | ~763.5s | |
#156 | GLM-4 | 32B | INT4 | 9 tok/s | ~34.8s | |
#157 | DeepSeek-R1 70B | 70B | INT4 | 5 tok/s | ~88.8s | |
#158 | GLM-4V | 9B | INT4 Q8 FP16 | 28 tok/s 19 tok/s 11 tok/s | ~10.0s ~10.0s ~10.1s | |
#159 | Gemma 3 27B | 27B | INT4 Q8 | 14 tok/s 8 tok/s | ~29.4s ~32.0s | |
#160 | Phi-4 | 14B | INT4 Q8 FP16 | 26 tok/s 15 tok/s 8 tok/s | ~15.5s ~15.5s ~16.9s | |
#161 | Llama 3.3 70B | 70B | INT4 | 5 tok/s | ~88.8s | |
#162 | DeepSeek-R1 32B | 32B | INT4 Q8 | 13 tok/s 7 tok/s | ~34.7s ~37.8s | |
#163 | Hunyuan Standard | 52B (389B active) | INT4 | 1 tok/s | ~477.4s | |
#164 | Qwen2.5-32B | 32B | INT4 Q8 | 12 tok/s 7 tok/s | ~34.7s ~37.8s | |
#168 | Magistral Small | 24B | INT4 Q8 | 12 tok/s 7 tok/s | ~26.2s ~28.6s | |
#169 | OLMo 3 7B Think | 7B | INT4 Q8 FP16 | 34 tok/s 23 tok/s 14 tok/s | ~7.8s ~7.9s ~7.9s | |
#170 | OLMo 3.1 32B Instruct | 32B | INT4 | 9 tok/s | ~34.8s | |
#171 | Llama 3.1 70B | 70B | INT4 | 5 tok/s | ~88.8s | |
#172 | DeepSeek-R1 14B | 14B | INT4 Q8 FP16 | 27 tok/s 16 tok/s 8 tok/s | ~15.5s ~15.5s ~16.9s | |
#173 | Ministral 3 14B | 14B | INT4 Q8 FP16 | 19 tok/s 13 tok/s 7 tok/s | ~15.5s ~15.5s ~16.9s | |
#177 | Gemma 3 12B | 12B | INT4 Q8 FP16 | 30 tok/s 18 tok/s 9 tok/s | ~13.3s ~13.3s ~13.4s | |
#180 | OLMo 3 32B Base | 32B | INT4 | 9 tok/s | ~34.8s | |
#181 | Mistral-Small-2501 | 24B | INT4 Q8 | 16 tok/s 9 tok/s | ~26.2s ~26.3s | |
#182 | Ministral 3 8B | 8B | INT4 Q8 FP16 | 30 tok/s 21 tok/s 12 tok/s | ~8.9s ~8.9s ~9.0s | |
#183 | Qwen3.5-2B | 2B | INT4 Q8 FP16 | 114 tok/s 79 tok/s 47 tok/s | ~2.4s ~2.4s ~2.4s | |
#184 | MiniCPM5-2B | 2B | INT4 Q8 FP16 | 114 tok/s 79 tok/s 47 tok/s | ~2.4s ~2.4s ~2.4s | |
#185 | Qwen2.5-14B | 14B | INT4 Q8 FP16 | 26 tok/s 15 tok/s 8 tok/s | ~15.5s ~15.5s ~16.9s | |
#186 | Gemma 2 27B | 27B | INT4 Q8 | 14 tok/s 8 tok/s | ~29.4s ~32.0s | |
#187 | OLMo 3.1 32B Think | 32B | INT4 | 9 tok/s | ~34.8s | |
#188 | Gemma 4 E2B | 5.1B | INT4 Q8 FP16 | 61 tok/s 38 tok/s 21 tok/s | ~5.8s ~5.8s ~5.8s | |
#189 | Qwen3-4B | 4B | INT4 Q8 FP16 | 73 tok/s 47 tok/s 26 tok/s | ~4.6s ~4.6s ~4.6s | |
#190 | DeepSeek-R1 8B | 8B | INT4 Q8 FP16 | 44 tok/s 26 tok/s 14 tok/s | ~8.9s ~8.9s ~9.0s | |
#191 | Llama 3 70B | 70B | INT4 | 5 tok/s | ~88.8s | |
#193 | Ministral 3 3B | 3B | INT4 Q8 FP16 | 58 tok/s 44 tok/s 28 tok/s | ~3.5s ~3.5s ~3.5s | |
#195 | Qwen3-14B | 14B | INT4 Q8 FP16 | 26 tok/s 15 tok/s 8 tok/s | ~15.5s ~15.5s ~16.9s | |
#196 | Kimi Linear 48B A3B Instruct | 48B (3B active) | INT4 | 22 tok/s | ~5.4s | |
#197 | Gemma 2 9B | 9B | INT4 Q8 FP16 | 39 tok/s 23 tok/s 12 tok/s | ~10.0s ~10.0s ~10.1s | |
#198 | Qwen2.5-7B | 7B | INT4 Q8 FP16 | 48 tok/s 29 tok/s 15 tok/s | ~7.8s ~7.8s ~7.9s | |
#199 | Qwen3-8B | 8B | INT4 Q8 FP16 | 43 tok/s 26 tok/s 14 tok/s | ~8.9s ~8.9s ~9.0s | |
#200 | Gemma 3 4B | 4B | INT4 Q8 FP16 | 73 tok/s 47 tok/s 26 tok/s | ~4.6s ~4.6s ~4.6s | |
#201 | DeepSeek-R1 7B | 7B | INT4 Q8 FP16 | 49 tok/s 29 tok/s 16 tok/s | ~7.8s ~7.8s ~7.9s | |
#202 | Llama 3.1 8B | 8B | INT4 Q8 FP16 | 43 tok/s 26 tok/s 14 tok/s | ~8.9s ~8.9s ~9.0s | |
#204 | Qwen3-1.7B | 1.7B | INT4 Q8 FP16 | 125 tok/s 89 tok/s 54 tok/s | ~2.1s ~2.1s ~2.1s | |
#205 | Command R | 35B | INT4 | 9 tok/s | ~37.9s | |
#206 | Gemma 3 270M | 0.27B | INT4 Q8 FP16 | 219 tok/s 198 tok/s 161 tok/s | ~325 ms ~325 ms ~326 ms | |
#208 | Llama 3 8B | 8B | INT4 Q8 FP16 | 43 tok/s 26 tok/s 14 tok/s | ~8.9s ~8.9s ~9.0s | |
#209 | Qwen3-0.6B | 0.6B | INT4 Q8 FP16 | 203 tok/s 164 tok/s 115 tok/s | ~729 ms ~730 ms ~733 ms | |
#210 | Phi-4-Mini | 3.8B | INT4 Q8 FP16 | 76 tok/s 49 tok/s 27 tok/s | ~4.4s ~4.4s ~4.4s | |
#211 | Mixtral-8x7B-v0.1 | 46.7B (7B active) | INT4 | 42 tok/s | ~9.1s | |
#212 | Ministral-8B-2410 | 8B | INT4 Q8 FP16 | 43 tok/s 26 tok/s 14 tok/s | ~8.9s ~8.9s ~9.0s | |
#213 | DeepSeek-R1 1.5B | 1.5B | INT4 Q8 FP16 | 135 tok/s 97 tok/s 60 tok/s | ~1.8s ~1.8s ~1.8s | |
#214 | Llama 3.2 3B | 3B | INT4 Q8 FP16 | 89 tok/s 59 tok/s 33 tok/s | ~3.5s ~3.5s ~3.5s | |
#215 | Qwen3.5-0.8B | 0.8B | INT4 Q8 FP16 | 182 tok/s 142 tok/s 95 tok/s | ~987 ms ~988 ms ~992 ms | |
#216 | Phi-3-mini | 3.8B | INT4 Q8 FP16 | 76 tok/s 49 tok/s 27 tok/s | ~4.4s ~4.4s ~4.4s | |
#217 | Gemma 3n E2B IT | 6B (2B active) | INT4 Q8 FP16 | 58 tok/s 49 tok/s 34 tok/s | ~2.6s ~2.6s ~2.6s | |
#218 | Mistral-7B-Instruct-v0.1 | 7.3B | INT4 Q8 FP16 | 46 tok/s 28 tok/s 15 tok/s | ~8.2s ~8.2s ~8.2s | |
#219 | Llama 3.2 1B | 1B | INT4 Q8 FP16 | 164 tok/s 124 tok/s 81 tok/s | ~1.3s ~1.3s ~1.3s | |
#220 | Gemma 2 2B | 2B | INT4 Q8 FP16 | 114 tok/s 79 tok/s 47 tok/s | ~2.4s ~2.4s ~2.4s | |
#221 | Phi-3-medium | 14B | INT4 Q8 FP16 | 26 tok/s 15 tok/s 8 tok/s | ~15.5s ~15.5s ~16.9s | |
#222 | OLMo 3 7B Instruct | 7B | INT4 Q8 FP16 | 34 tok/s 23 tok/s 14 tok/s | ~7.8s ~7.9s ~7.9s | |
#223 | Yi-34B | 34B | INT4 | 9 tok/s | ~36.9s | |
#224 | Qwen2-7B | 7B | INT4 Q8 FP16 | 48 tok/s 29 tok/s 15 tok/s | ~7.8s ~7.8s ~7.9s | |
#225 | Phi-3-small | 7B | INT4 Q8 FP16 | 48 tok/s 29 tok/s 15 tok/s | ~7.8s ~7.8s ~7.9s | |
#226 | Gemma 3 1B | 1B | INT4 Q8 FP16 | 164 tok/s 124 tok/s 81 tok/s | ~1.3s ~1.3s ~1.3s | |
#227 | Mistral-7B-Instruct-v0.2 | 7.3B | INT4 Q8 FP16 | 46 tok/s 28 tok/s 15 tok/s | ~8.2s ~8.2s ~8.2s | |
#229 | Gemma 1 7B | 7B | INT4 Q8 FP16 | 34 tok/s 23 tok/s 14 tok/s | ~7.8s ~7.9s ~7.9s | |
#230 | Gemma 1 2B | 2B | INT4 Q8 FP16 | 122 tok/s 83 tok/s 48 tok/s | ~2.4s ~2.4s ~2.4s | |
#231 | ChatGLM3-6B | 6B | INT4 Q8 FP16 | 38 tok/s 27 tok/s 16 tok/s | ~6.8s ~6.8s ~6.8s | |
#232 | ChatGLM2-6B | 6B | INT4 Q8 FP16 | 38 tok/s 27 tok/s 16 tok/s | ~6.8s ~6.8s ~6.8s | |
#233 | ChatGLM-6B | 6B | INT4 Q8 FP16 | 38 tok/s 27 tok/s 16 tok/s | ~6.8s ~6.8s ~6.8s | |
未上榜模型（按推理速度排序） | ||||||
- | Qwen2-0.5B | 0.5B | INT4 Q8 FP16 | 215 tok/s 178 tok/s 129 tok/s | ~604 ms ~605 ms ~607 ms | |
- | Qwen2.5-0.5B | 0.5B | INT4 Q8 FP16 | 215 tok/s 178 tok/s 129 tok/s | ~604 ms ~605 ms ~607 ms | |
- | ERNIE-4.5-0.3B | 0.3B | INT4 Q8 FP16 | 211 tok/s 189 tok/s 152 tok/s | ~362 ms ~362 ms ~364 ms | |
- | ERNIE-4.5-0.3B-Base | 0.3B | INT4 Q8 FP16 | 211 tok/s 189 tok/s 152 tok/s | ~362 ms ~362 ms ~364 ms | |
- | Falcon-1B | 1B | INT4 Q8 FP16 | 174 tok/s 130 tok/s 83 tok/s | ~1.3s ~1.3s ~1.3s | |
- | Falcon3-1B | 1B | INT4 Q8 FP16 | 164 tok/s 124 tok/s 81 tok/s | ~1.3s ~1.3s ~1.3s | |
- | Qwen2-1.5B | 1.5B | INT4 Q8 FP16 | 135 tok/s 97 tok/s 60 tok/s | ~1.8s ~1.8s ~1.8s | |
- | Qwen2.5-1.5B | 1.5B | INT4 Q8 FP16 | 135 tok/s 97 tok/s 60 tok/s | ~1.8s ~1.8s ~1.8s | |
- | Spark X2.5 1.7B | 1.7B | INT4 Q8 FP16 | 125 tok/s 89 tok/s 54 tok/s | ~2.1s ~2.1s ~2.1s | |
- | Hy-MT2 1.8B | 1.8B | INT4 Q8 FP16 | 122 tok/s 85 tok/s 51 tok/s | ~2.2s ~2.2s ~2.2s | |
- | Falcon-3B | 3B | INT4 Q8 FP16 | 95 tok/s 61 tok/s 34 tok/s | ~3.5s ~3.5s ~3.5s | |
- | CroissantLLM Base | 1.3B | INT4 Q8 FP16 | 93 tok/s 77 tok/s 54 tok/s | ~1.6s ~1.6s ~1.6s | |
- | Phi-1 | 1.3B | INT4 Q8 FP16 | 93 tok/s 77 tok/s 54 tok/s | ~1.6s ~1.6s ~1.6s | |
- | Phi-1.5 | 1.3B | INT4 Q8 FP16 | 93 tok/s 77 tok/s 54 tok/s | ~1.6s ~1.6s ~1.6s | |
- | DeepSeek-R1 3B | 3B | INT4 Q8 FP16 | 91 tok/s 59 tok/s 34 tok/s | ~3.5s ~3.5s ~3.5s | |
- | Falcon3-3B | 3B | INT4 Q8 FP16 | 89 tok/s 59 tok/s 33 tok/s | ~3.5s ~3.5s ~3.5s | |
- | Qwen2.5-3B | 3B | INT4 Q8 FP16 | 89 tok/s 59 tok/s 33 tok/s | ~3.5s ~3.5s ~3.5s | |
- | ERNIE-4.5-21B-A3B | 21B (3B active) | INT4 Q8 | 77 tok/s 54 tok/s | ~4.3s ~4.3s | |
- | ERNIE-4.5-21B-A3B-Base | 21B (3B active) | INT4 Q8 | 77 tok/s 54 tok/s | ~4.3s ~4.3s | |
- | ERNIE-4.5-VL-28B-A3B | 28B (3B active) | INT4 Q8 | 74 tok/s 50 tok/s | ~4.5s ~4.9s | |
- | ERNIE-4.5-VL-28B-A3B-Base | 28B (3B active) | INT4 Q8 | 74 tok/s 50 tok/s | ~4.5s ~4.9s | |
- | Hy-MT2 30B A3B | 30B (3B active) | INT4 Q8 | 73 tok/s 49 tok/s | ~4.6s ~5.0s | |
- | Spark X2.5 4B | 4B | INT4 Q8 FP16 | 73 tok/s 47 tok/s 26 tok/s | ~4.6s ~4.6s ~4.6s | |
- | Phi-2 | 2.7B | INT4 Q8 FP16 | 61 tok/s 47 tok/s 31 tok/s | ~3.2s ~3.2s ~3.2s | |
- | MaLLaM-3B | 3B | INT4 Q8 FP16 | 58 tok/s 44 tok/s 28 tok/s | ~3.5s ~3.5s ~3.5s | |
- | SmolLM3 3B | 3B | INT4 Q8 FP16 | 58 tok/s 44 tok/s 28 tok/s | ~3.5s ~3.5s ~3.5s | |
- | Falcon-7B | 7B | INT4 Q8 FP16 | 50 tok/s 30 tok/s 16 tok/s | ~7.8s ~7.8s ~7.9s | |
- | Falcon3-7B | 7B | INT4 Q8 FP16 | 48 tok/s 29 tok/s 15 tok/s | ~7.8s ~7.8s ~7.9s | |
- | Hy-MT2 7B | 7B | INT4 Q8 FP16 | 48 tok/s 29 tok/s 15 tok/s | ~7.8s ~7.8s ~7.9s | |
- | Mistral-7B-v0.1 | 7.3B | INT4 Q8 FP16 | 46 tok/s 28 tok/s 15 tok/s | ~8.2s ~8.2s ~8.2s | |
- | LensVLM 9B | 9B | INT4 Q8 FP16 | 39 tok/s 23 tok/s 12 tok/s | ~10.0s ~10.0s ~10.1s | |
- | MiMo V2.6 Distill Qwen 9B | 9B | INT4 Q8 FP16 | 39 tok/s 23 tok/s 12 tok/s | ~10.0s ~10.0s ~10.1s | |
- | ChatGLM3-6B-32K | 6B | INT4 Q8 FP16 | 38 tok/s 27 tok/s 16 tok/s | ~6.8s ~6.8s ~6.8s | |
- | Yi-6B | 6B | INT4 Q8 FP16 | 38 tok/s 27 tok/s 16 tok/s | ~6.8s ~6.8s ~6.8s | |
- | Falcon3-10B | 10B | INT4 Q8 FP16 | 35 tok/s 21 tok/s 11 tok/s | ~11.1s ~11.1s ~11.2s | |
- | Kimi-VL-A3B-Instruct | 16B (3B active) | INT4 Q8 FP16 | 35 tok/s 31 tok/s 20 tok/s | ~4.1s ~4.1s ~4.9s | |
- | Falcon2-11B | 11B | INT4 Q8 FP16 | 34 tok/s 20 tok/s 10 tok/s | ~12.2s ~12.2s ~12.2s | |
- | Hunyuan Lite | 7B | INT4 Q8 FP16 | 34 tok/s 23 tok/s 14 tok/s | ~7.8s ~7.9s ~7.9s | |
- | OLMo 3 7B Base | 7B | INT4 Q8 FP16 | 34 tok/s 23 tok/s 14 tok/s | ~7.8s ~7.9s ~7.9s | |
- | SEA-LION-7B | 7.1B | INT4 Q8 FP16 | 33 tok/s 23 tok/s 14 tok/s | ~7.9s ~8.0s ~8.0s | |
- | SEA-LION-7B-Instruct | 7.1B | INT4 Q8 FP16 | 33 tok/s 23 tok/s 14 tok/s | ~7.9s ~8.0s ~8.0s | |
- | Sahabat-AI-Llama3-8B-Instruct | 8B | INT4 Q8 FP16 | 30 tok/s 21 tok/s 12 tok/s | ~8.9s ~8.9s ~9.0s | |
- | Typhoon-2-8B | 8B | INT4 Q8 FP16 | 30 tok/s 21 tok/s 12 tok/s | ~8.9s ~8.9s ~9.0s | |
- | GLM-4-9B | 9B | INT4 Q8 FP16 | 28 tok/s 19 tok/s 11 tok/s | ~10.0s ~10.0s ~10.1s | |
- | GLM-4-9B-Chat | 9B | INT4 Q8 FP16 | 28 tok/s 19 tok/s 11 tok/s | ~10.0s ~10.0s ~10.1s | |
- | GLM-4-9B-Chat-1M | 9B | INT4 Q8 FP16 | 28 tok/s 19 tok/s 11 tok/s | ~10.0s ~10.0s ~10.1s | |
- | Yi-9B | 9B | INT4 Q8 FP16 | 28 tok/s 19 tok/s 11 tok/s | ~10.0s ~10.0s ~10.1s | |
- | Sahabat-AI-Gemma2-9B | 9.2B | INT4 Q8 FP16 | 27 tok/s 19 tok/s 11 tok/s | ~10.2s ~10.3s ~10.3s | |
- | Sahabat-AI-Gemma2-9B-Instruct | 9.2B | INT4 Q8 FP16 | 27 tok/s 19 tok/s 11 tok/s | ~10.2s ~10.3s ~10.3s | |
- | GreenMind-14B-R1 | 14B | INT4 Q8 FP16 | 19 tok/s 13 tok/s 7 tok/s | ~15.5s ~15.5s ~16.9s | |
- | Falcon-40B | 40B | INT4 | 10 tok/s | ~43.2s | |
推理能力评估
不同量化精度下可运行的最大模型参数级别。
上下文长度:
4位量化
32B 模型
8位精度
14B 模型
16位全精度
14B 模型
微调能力评估
本地模型训练与适配器微调支持能力。
微调上下文:
QLoRA
32B 模型
LoRA
14B 模型
全参数微调
3B 模型
负载与选型建议
Apple M4 Pro (48GB) 上的最佳量化精度与运行规格建议。
最佳推理区间
适合在全精度或 Q8 精度下运行 14B 模型，以及在 Q4 精度下搭配最高 32k 上下文运行 32B 模型。
显存带宽与推理速率
具备 273 GB/s 聚合带宽，单批次推理处于显存带宽受限模式。8B Q4 模型预估生成速度约 61 tok/s，70B Q4 模型预估生成速度约 7 tok/s。
统一内存配额
采用 75% 可用分配预算建模，以保留系统内存供 macOS 系统及图形渲染进程稳定运行。