趋近智
Apple M3 Max (36GB)
Apple M3 Max (36GB) 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（27 GB）及 300 GB/s 统一内存带宽进行建模。
总显存
36 GB
可用显存上限
27 GB （75% 可用预算）
显存带宽
300 GB/s
最大推理模型级别
32B 模型 (Q4)
热设计功耗
115W
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 M3 Max (36GB) 最佳运行大模型与速度矩阵
主流开源大模型在 Apple M3 Max (36GB) 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 | 16 tok/s | ~25.4s | |
#48 | Agnes 3.0 Flash | 33B | INT4 | 13 tok/s | ~30.9s | |
#72 | K2 Horizon MoVA 36B A4B | 36B (4B active) | INT4 | 66 tok/s | ~5.0s | |
#83 | Qwen3.5-27B | 27B | INT4 | 16 tok/s | ~25.4s | |
#92 | Sarvam-30B | 32B (2.4B active) | INT4 | 87 tok/s | ~3.5s | |
#94 | Qwen3.6 35B A3B | 35B (3B active) | INT4 | 78 tok/s | ~4.0s | |
#96 | Muse Glimmer 30B | 30B | INT4 | 14 tok/s | ~28.2s | |
#100 | Phi-4 Reasoning Plus | 14B | INT4 Q8 | 21 tok/s 14 tok/s | ~13.4s ~13.4s | |
#101 | MiMo V2 Flash | 15B (309B active) | INT4 Q8 | 1 tok/s 1 tok/s | ~278.0s ~278.6s | |
#108 | Gemma 4 26B A4B | 25.2B (3.8B active) | INT4 Q8 | 71 tok/s 44 tok/s | ~4.5s ~5.4s | |
#109 | Gemma 4 12B | 11.95B | INT4 Q8 FP16 | 24 tok/s 16 tok/s 8 tok/s | ~11.4s ~11.5s ~13.5s | |
#113 | Qwen3.5-9B | 9B | INT4 Q8 FP16 | 42 tok/s 25 tok/s 13 tok/s | ~8.7s ~8.7s ~8.7s | |
#114 | Qwen3.5-35B-A3B | 35B (3B active) | INT4 | 78 tok/s | ~4.0s | |
#118 | K2 Horizon 7B | 7B | INT4 Q8 FP16 | 52 tok/s 31 tok/s 17 tok/s | ~6.8s ~6.8s ~6.8s | |
#128 | NVIDIA Nemotron 3 Nano 30B-A3B | 3.5B (30B active) | INT4 Q8 FP16 | 14 tok/s 8 tok/s 4 tok/s | ~27.3s ~27.3s ~27.4s | |
#129 | Gemma 4 31B | 30.7B | INT4 | 14 tok/s | ~28.8s | |
#131 | Nemotron 3.5 Lightning | 30B (3B active) | INT4 | 28 tok/s | ~4.3s | |
#132 | Qwen3.5-4B | 4B | INT4 Q8 FP16 | 79 tok/s 51 tok/s 28 tok/s | ~4.0s ~4.0s ~4.0s | |
#133 | Qwen3-30B-A3B | 30B (3B active) | INT4 | 79 tok/s | ~4.0s | |
#136 | GPT-OSS 20B | 21B (3.6B active) | INT4 Q8 | 33 tok/s 26 tok/s | ~4.3s ~5.0s | |
#146 | Granite 4.2 8B | 8B | INT4 Q8 FP16 | 46 tok/s 28 tok/s 15 tok/s | ~7.7s ~7.7s ~7.7s | |
#150 | Gemma 4 E4B | 8B | INT4 Q8 FP16 | 46 tok/s 28 tok/s 15 tok/s | ~7.7s ~7.7s ~7.7s | |
#151 | Nemotron 3.5 Lightning 30B A3B | 30B (3B active) | INT4 | 79 tok/s | ~4.0s | |
#153 | Qwen3-32B | 32B | INT4 | 13 tok/s | ~30.0s | |
#156 | GLM-4 | 32B | INT4 | 10 tok/s | ~32.6s | |
#158 | GLM-4V | 9B | INT4 Q8 FP16 | 30 tok/s 21 tok/s 12 tok/s | ~8.7s ~8.7s ~8.7s | |
#159 | Gemma 3 27B | 27B | INT4 | 16 tok/s | ~25.4s | |
#160 | Phi-4 | 14B | INT4 Q8 | 29 tok/s 17 tok/s | ~13.4s ~13.4s | |
#162 | DeepSeek-R1 32B | 32B | INT4 | 14 tok/s | ~30.0s | |
#164 | Qwen2.5-32B | 32B | INT4 | 13 tok/s | ~30.0s | |
#168 | Magistral Small | 24B | INT4 | 13 tok/s | ~22.7s | |
#169 | OLMo 3 7B Think | 7B | INT4 Q8 FP16 | 37 tok/s 26 tok/s 15 tok/s | ~6.8s ~6.8s ~6.8s | |
#170 | OLMo 3.1 32B Instruct | 32B | INT4 | 10 tok/s | ~32.6s | |
#172 | DeepSeek-R1 14B | 14B | INT4 Q8 | 29 tok/s 17 tok/s | ~13.4s ~13.4s | |
#173 | Ministral 3 14B | 14B | INT4 Q8 | 21 tok/s 14 tok/s | ~13.4s ~13.4s | |
#177 | Gemma 3 12B | 12B | INT4 Q8 FP16 | 33 tok/s 19 tok/s 10 tok/s | ~11.5s ~11.5s ~12.6s | |
#180 | OLMo 3 32B Base | 32B | INT4 | 10 tok/s | ~32.6s | |
#181 | Mistral-Small-2501 | 24B | INT4 Q8 | 18 tok/s 9 tok/s | ~22.6s ~26.7s | |
#182 | Ministral 3 8B | 8B | INT4 Q8 FP16 | 33 tok/s 23 tok/s 13 tok/s | ~7.7s ~7.7s ~7.8s | |
#183 | Qwen3.5-2B | 2B | INT4 Q8 FP16 | 121 tok/s 85 tok/s 51 tok/s | ~2.1s ~2.1s ~2.1s | |
#184 | MiniCPM5-2B | 2B | INT4 Q8 FP16 | 121 tok/s 85 tok/s 51 tok/s | ~2.1s ~2.1s ~2.1s | |
#185 | Qwen2.5-14B | 14B | INT4 Q8 | 29 tok/s 17 tok/s | ~13.4s ~13.4s | |
#186 | Gemma 2 27B | 27B | INT4 | 16 tok/s | ~25.4s | |
#187 | OLMo 3.1 32B Think | 32B | INT4 | 10 tok/s | ~32.6s | |
#188 | Gemma 4 E2B | 5.1B | INT4 Q8 FP16 | 66 tok/s 41 tok/s 23 tok/s | ~5.0s ~5.0s ~5.0s | |
#189 | Qwen3-4B | 4B | INT4 Q8 FP16 | 79 tok/s 51 tok/s 28 tok/s | ~4.0s ~4.0s ~4.0s | |
#190 | DeepSeek-R1 8B | 8B | INT4 Q8 FP16 | 48 tok/s 28 tok/s 15 tok/s | ~7.7s ~7.7s ~7.7s | |
#193 | Ministral 3 3B | 3B | INT4 Q8 FP16 | 63 tok/s 48 tok/s 31 tok/s | ~3.0s ~3.0s ~3.1s | |
#195 | Qwen3-14B | 14B | INT4 Q8 | 29 tok/s 17 tok/s | ~13.4s ~13.4s | |
#197 | Gemma 2 9B | 9B | INT4 Q8 FP16 | 42 tok/s 25 tok/s 13 tok/s | ~8.7s ~8.7s ~8.7s | |
#198 | Qwen2.5-7B | 7B | INT4 Q8 FP16 | 52 tok/s 31 tok/s 17 tok/s | ~6.8s ~6.8s ~6.8s | |
#199 | Qwen3-8B | 8B | INT4 Q8 FP16 | 46 tok/s 28 tok/s 15 tok/s | ~7.7s ~7.7s ~7.7s | |
#200 | Gemma 3 4B | 4B | INT4 Q8 FP16 | 79 tok/s 51 tok/s 28 tok/s | ~4.0s ~4.0s ~4.0s | |
#201 | DeepSeek-R1 7B | 7B | INT4 Q8 FP16 | 53 tok/s 32 tok/s 17 tok/s | ~6.8s ~6.8s ~6.8s | |
#202 | Llama 3.1 8B | 8B | INT4 Q8 FP16 | 46 tok/s 28 tok/s 15 tok/s | ~7.7s ~7.7s ~7.7s | |
#204 | Qwen3-1.7B | 1.7B | INT4 Q8 FP16 | 132 tok/s 95 tok/s 58 tok/s | ~1.8s ~1.8s ~1.8s | |
#205 | Command R | 35B | INT4 | 9 tok/s | ~35.6s | |
#206 | Gemma 3 270M | 0.27B | INT4 Q8 FP16 | 224 tok/s 204 tok/s 168 tok/s | ~282 ms ~282 ms ~283 ms | |
#208 | Llama 3 8B | 8B | INT4 Q8 FP16 | 46 tok/s 28 tok/s 15 tok/s | ~7.7s ~7.7s ~7.7s | |
#209 | Qwen3-0.6B | 0.6B | INT4 Q8 FP16 | 209 tok/s 171 tok/s 122 tok/s | ~631 ms ~632 ms ~634 ms | |
#210 | Phi-4-Mini | 3.8B | INT4 Q8 FP16 | 82 tok/s 53 tok/s 30 tok/s | ~3.8s ~3.8s ~3.8s | |
#211 | Mixtral-8x7B-v0.1 | 46.7B (7B active) | INT4 | 41 tok/s | ~9.2s | |
#212 | Ministral-8B-2410 | 8B | INT4 Q8 FP16 | 46 tok/s 28 tok/s 15 tok/s | ~7.7s ~7.7s ~7.7s | |
#213 | DeepSeek-R1 1.5B | 1.5B | INT4 Q8 FP16 | 142 tok/s 103 tok/s 64 tok/s | ~1.6s ~1.6s ~1.6s | |
#214 | Llama 3.2 3B | 3B | INT4 Q8 FP16 | 95 tok/s 63 tok/s 36 tok/s | ~3.0s ~3.0s ~3.0s | |
#215 | Qwen3.5-0.8B | 0.8B | INT4 Q8 FP16 | 188 tok/s 149 tok/s 101 tok/s | ~854 ms ~855 ms ~858 ms | |
#216 | Phi-3-mini | 3.8B | INT4 Q8 FP16 | 82 tok/s 53 tok/s 30 tok/s | ~3.8s ~3.8s ~3.8s | |
#217 | Gemma 3n E2B IT | 6B (2B active) | INT4 Q8 FP16 | 62 tok/s 53 tok/s 37 tok/s | ~2.3s ~2.3s ~2.3s | |
#218 | Mistral-7B-Instruct-v0.1 | 7.3B | INT4 Q8 FP16 | 50 tok/s 30 tok/s 16 tok/s | ~7.0s ~7.1s ~7.1s | |
#219 | Llama 3.2 1B | 1B | INT4 Q8 FP16 | 171 tok/s 131 tok/s 87 tok/s | ~1.1s ~1.1s ~1.1s | |
#220 | Gemma 2 2B | 2B | INT4 Q8 FP16 | 121 tok/s 85 tok/s 51 tok/s | ~2.1s ~2.1s ~2.1s | |
#221 | Phi-3-medium | 14B | INT4 Q8 | 29 tok/s 17 tok/s | ~13.4s ~13.4s | |
#222 | OLMo 3 7B Instruct | 7B | INT4 Q8 FP16 | 37 tok/s 26 tok/s 15 tok/s | ~6.8s ~6.8s ~6.8s | |
#223 | Yi-34B | 34B | INT4 | 9 tok/s | ~34.6s | |
#224 | Qwen2-7B | 7B | INT4 Q8 FP16 | 52 tok/s 31 tok/s 17 tok/s | ~6.8s ~6.8s ~6.8s | |
#225 | Phi-3-small | 7B | INT4 Q8 FP16 | 52 tok/s 31 tok/s 17 tok/s | ~6.8s ~6.8s ~6.8s | |
#226 | Gemma 3 1B | 1B | INT4 Q8 FP16 | 171 tok/s 131 tok/s 87 tok/s | ~1.1s ~1.1s ~1.1s | |
#227 | Mistral-7B-Instruct-v0.2 | 7.3B | INT4 Q8 FP16 | 50 tok/s 30 tok/s 16 tok/s | ~7.0s ~7.1s ~7.1s | |
#229 | Gemma 1 7B | 7B | INT4 Q8 FP16 | 37 tok/s 26 tok/s 15 tok/s | ~6.8s ~6.8s ~6.8s | |
#230 | Gemma 1 2B | 2B | INT4 Q8 FP16 | 129 tok/s 88 tok/s 52 tok/s | ~2.1s ~2.1s ~2.1s | |
#231 | ChatGLM3-6B | 6B | INT4 Q8 FP16 | 41 tok/s 29 tok/s 17 tok/s | ~5.8s ~5.9s ~5.9s | |
#232 | ChatGLM2-6B | 6B | INT4 Q8 FP16 | 41 tok/s 29 tok/s 17 tok/s | ~5.8s ~5.9s ~5.9s | |
#233 | ChatGLM-6B | 6B | INT4 Q8 FP16 | 41 tok/s 29 tok/s 17 tok/s | ~5.8s ~5.9s ~5.9s | |
未上榜模型（按推理速度排序） | ||||||
- | Qwen2-0.5B | 0.5B | INT4 Q8 FP16 | 221 tok/s 185 tok/s 136 tok/s | ~523 ms ~524 ms ~526 ms | |
- | Qwen2.5-0.5B | 0.5B | INT4 Q8 FP16 | 221 tok/s 185 tok/s 136 tok/s | ~523 ms ~524 ms ~526 ms | |
- | ERNIE-4.5-0.3B | 0.3B | INT4 Q8 FP16 | 217 tok/s 196 tok/s 159 tok/s | ~314 ms ~314 ms ~315 ms | |
- | ERNIE-4.5-0.3B-Base | 0.3B | INT4 Q8 FP16 | 217 tok/s 196 tok/s 159 tok/s | ~314 ms ~314 ms ~315 ms | |
- | Falcon-1B | 1B | INT4 Q8 FP16 | 181 tok/s 137 tok/s 89 tok/s | ~1.1s ~1.1s ~1.1s | |
- | Falcon3-1B | 1B | INT4 Q8 FP16 | 171 tok/s 131 tok/s 87 tok/s | ~1.1s ~1.1s ~1.1s | |
- | Qwen2-1.5B | 1.5B | INT4 Q8 FP16 | 142 tok/s 103 tok/s 64 tok/s | ~1.6s ~1.6s ~1.6s | |
- | Qwen2.5-1.5B | 1.5B | INT4 Q8 FP16 | 142 tok/s 103 tok/s 64 tok/s | ~1.6s ~1.6s ~1.6s | |
- | Spark X2.5 1.7B | 1.7B | INT4 Q8 FP16 | 132 tok/s 95 tok/s 58 tok/s | ~1.8s ~1.8s ~1.8s | |
- | Hy-MT2 1.8B | 1.8B | INT4 Q8 FP16 | 129 tok/s 91 tok/s 56 tok/s | ~1.9s ~1.9s ~1.9s | |
- | Falcon-3B | 3B | INT4 Q8 FP16 | 101 tok/s 66 tok/s 37 tok/s | ~3.0s ~3.0s ~3.0s | |
- | CroissantLLM Base | 1.3B | INT4 Q8 FP16 | 99 tok/s 83 tok/s 58 tok/s | ~1.4s ~1.4s ~1.4s | |
- | Phi-1 | 1.3B | INT4 Q8 FP16 | 99 tok/s 83 tok/s 58 tok/s | ~1.4s ~1.4s ~1.4s | |
- | Phi-1.5 | 1.3B | INT4 Q8 FP16 | 99 tok/s 83 tok/s 58 tok/s | ~1.4s ~1.4s ~1.4s | |
- | DeepSeek-R1 3B | 3B | INT4 Q8 FP16 | 97 tok/s 64 tok/s 36 tok/s | ~3.0s ~3.0s ~3.0s | |
- | Falcon3-3B | 3B | INT4 Q8 FP16 | 95 tok/s 63 tok/s 36 tok/s | ~3.0s ~3.0s ~3.0s | |
- | Qwen2.5-3B | 3B | INT4 Q8 FP16 | 95 tok/s 63 tok/s 36 tok/s | ~3.0s ~3.0s ~3.0s | |
- | ERNIE-4.5-21B-A3B | 21B (3B active) | INT4 Q8 | 83 tok/s 55 tok/s | ~3.7s ~4.0s | |
- | ERNIE-4.5-21B-A3B-Base | 21B (3B active) | INT4 Q8 | 83 tok/s 55 tok/s | ~3.7s ~4.0s | |
- | ERNIE-4.5-VL-28B-A3B | 28B (3B active) | INT4 | 80 tok/s | ~3.9s | |
- | ERNIE-4.5-VL-28B-A3B-Base | 28B (3B active) | INT4 | 80 tok/s | ~3.9s | |
- | Hy-MT2 30B A3B | 30B (3B active) | INT4 | 79 tok/s | ~4.0s | |
- | Spark X2.5 4B | 4B | INT4 Q8 FP16 | 79 tok/s 51 tok/s 28 tok/s | ~4.0s ~4.0s ~4.0s | |
- | Phi-2 | 2.7B | INT4 Q8 FP16 | 66 tok/s 51 tok/s 33 tok/s | ~2.8s ~2.8s ~2.8s | |
- | MaLLaM-3B | 3B | INT4 Q8 FP16 | 63 tok/s 48 tok/s 31 tok/s | ~3.0s ~3.0s ~3.1s | |
- | SmolLM3 3B | 3B | INT4 Q8 FP16 | 63 tok/s 48 tok/s 31 tok/s | ~3.0s ~3.0s ~3.1s | |
- | Falcon-7B | 7B | INT4 Q8 FP16 | 54 tok/s 32 tok/s 17 tok/s | ~6.8s ~6.8s ~6.8s | |
- | Falcon3-7B | 7B | INT4 Q8 FP16 | 52 tok/s 31 tok/s 17 tok/s | ~6.8s ~6.8s ~6.8s | |
- | Hy-MT2 7B | 7B | INT4 Q8 FP16 | 52 tok/s 31 tok/s 17 tok/s | ~6.8s ~6.8s ~6.8s | |
- | Mistral-7B-v0.1 | 7.3B | INT4 Q8 FP16 | 50 tok/s 30 tok/s 16 tok/s | ~7.0s ~7.1s ~7.1s | |
- | LensVLM 9B | 9B | INT4 Q8 FP16 | 42 tok/s 25 tok/s 13 tok/s | ~8.7s ~8.7s ~8.7s | |
- | MiMo V2.6 Distill Qwen 9B | 9B | INT4 Q8 FP16 | 42 tok/s 25 tok/s 13 tok/s | ~8.7s ~8.7s ~8.7s | |
- | ChatGLM3-6B-32K | 6B | INT4 Q8 FP16 | 41 tok/s 29 tok/s 17 tok/s | ~5.8s ~5.9s ~5.9s | |
- | Yi-6B | 6B | INT4 Q8 FP16 | 41 tok/s 29 tok/s 17 tok/s | ~5.8s ~5.9s ~5.9s | |
- | Falcon3-10B | 10B | INT4 Q8 FP16 | 38 tok/s 23 tok/s 12 tok/s | ~9.6s ~9.6s ~9.6s | |
- | Kimi-VL-A3B-Instruct | 16B (3B active) | INT4 Q8 | 38 tok/s 33 tok/s | ~3.6s ~3.6s | |
- | Falcon2-11B | 11B | INT4 Q8 FP16 | 37 tok/s 21 tok/s 11 tok/s | ~10.5s ~10.5s ~11.5s | |
- | Hunyuan Lite | 7B | INT4 Q8 FP16 | 37 tok/s 26 tok/s 15 tok/s | ~6.8s ~6.8s ~6.8s | |
- | OLMo 3 7B Base | 7B | INT4 Q8 FP16 | 37 tok/s 26 tok/s 15 tok/s | ~6.8s ~6.8s ~6.8s | |
- | SEA-LION-7B | 7.1B | INT4 Q8 FP16 | 36 tok/s 25 tok/s 15 tok/s | ~6.9s ~6.9s ~6.9s | |
- | SEA-LION-7B-Instruct | 7.1B | INT4 Q8 FP16 | 36 tok/s 25 tok/s 15 tok/s | ~6.9s ~6.9s ~6.9s | |
- | Sahabat-AI-Llama3-8B-Instruct | 8B | INT4 Q8 FP16 | 33 tok/s 23 tok/s 13 tok/s | ~7.7s ~7.7s ~7.8s | |
- | Typhoon-2-8B | 8B | INT4 Q8 FP16 | 33 tok/s 23 tok/s 13 tok/s | ~7.7s ~7.7s ~7.8s | |
- | GLM-4-9B | 9B | INT4 Q8 FP16 | 30 tok/s 21 tok/s 12 tok/s | ~8.7s ~8.7s ~8.7s | |
- | GLM-4-9B-Chat | 9B | INT4 Q8 FP16 | 30 tok/s 21 tok/s 12 tok/s | ~8.7s ~8.7s ~8.7s | |
- | GLM-4-9B-Chat-1M | 9B | INT4 Q8 FP16 | 30 tok/s 21 tok/s 12 tok/s | ~8.7s ~8.7s ~8.7s | |
- | Sahabat-AI-Gemma2-9B | 9.2B | INT4 Q8 FP16 | 30 tok/s 20 tok/s 11 tok/s | ~8.8s ~8.9s ~9.7s | |
- | Sahabat-AI-Gemma2-9B-Instruct | 9.2B | INT4 Q8 FP16 | 30 tok/s 20 tok/s 11 tok/s | ~8.8s ~8.9s ~9.7s | |
- | Yi-9B | 9B | INT4 Q8 FP16 | 30 tok/s 21 tok/s 12 tok/s | ~8.7s ~8.7s ~8.7s | |
- | GreenMind-14B-R1 | 14B | INT4 Q8 | 21 tok/s 14 tok/s | ~13.4s ~13.4s | |
- | Falcon-40B | 40B | INT4 | 11 tok/s | ~40.6s | |
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
8B 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 M3 Max (36GB) 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 300 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 67 tok/s，70B Q4 模型预估生成速度约 8 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。