趋近智
AMD RX 6700 (10GB)
AMD RX 6700 (10GB) 提供 10 GB 显存/内存与 320 GB/s 带宽，支持通过 ROCm 与 Vulkan 加速进行本地推理。
总显存
10 GB
可用显存上限
10 GB
显存带宽
320 GB/s
最大推理模型级别
8B 模型 (Q4)
热设计功耗
175W
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:
AMD RX 6700 (10GB) 最佳运行大模型与速度矩阵
主流开源大模型在 AMD RX 6700 (10GB) 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:
显卡数量:
| 排名 | 模型 | 参数量 | 量化精度 | 预估 TPS | TTFT | 可运行 |
|---|---|---|---|---|---|---|
#124 | Qwen3.5-9B | 9B | INT4 | 44 tok/s | ~5.0s | |
#129 | K2 Horizon 7B | 7B | INT4 Q8 | 54 tok/s 31 tok/s | ~3.9s ~4.3s | |
#141 | Qwen3.5-4B | 4B | INT4 Q8 FP16 | 81 tok/s 53 tok/s 28 tok/s | ~2.3s ~2.3s ~2.5s | |
#146 | GLM-4V | 9B | INT4 | 30 tok/s | ~5.5s | |
#151 | NVIDIA Nemotron 3 Nano 30B-A3B | 3.5B (30B active) | INT4 Q8 FP16 | 14 tok/s 8 tok/s 4 tok/s | ~15.8s ~15.9s ~18.8s | |
#159 | Granite 4.2 8B | 8B | INT4 Q8 | 48 tok/s 26 tok/s | ~4.5s ~5.3s | |
#160 | Gemma 4 E4B | 8B | INT4 Q8 | 48 tok/s 26 tok/s | ~4.5s ~5.3s | |
#167 | Phi-4 | 14B | INT4 | 29 tok/s | ~8.4s | |
#179 | DeepSeek-R1 14B | 14B | INT4 | 29 tok/s | ~8.4s | |
#180 | OLMo 3 7B Think | 7B | INT4 Q8 | 38 tok/s 24 tok/s | ~3.9s ~4.6s | |
#187 | Gemma 3 12B | 12B | INT4 | 34 tok/s | ~6.7s | |
#190 | Ministral 3 8B | 8B | INT4 | 35 tok/s | ~4.5s | |
#192 | MiniCPM5-2B | 2B | INT4 Q8 FP16 | 122 tok/s 87 tok/s 53 tok/s | ~1.2s ~1.2s ~1.2s | |
#193 | Qwen3.5-2B | 2B | INT4 Q8 FP16 | 122 tok/s 87 tok/s 53 tok/s | ~1.2s ~1.2s ~1.2s | |
#195 | Qwen2.5-14B | 14B | INT4 | 29 tok/s | ~8.4s | |
#200 | Gemma 4 E2B | 5.1B | INT4 Q8 | 68 tok/s 43 tok/s | ~2.9s ~2.9s | |
#203 | DeepSeek-R1 8B | 8B | INT4 Q8 | 50 tok/s 28 tok/s | ~4.5s ~4.9s | |
#205 | Qwen3-4B | 4B | INT4 Q8 FP16 | 81 tok/s 53 tok/s 28 tok/s | ~2.3s ~2.3s ~2.5s | |
#206 | Ministral 3 3B | 3B | INT4 Q8 FP16 | 65 tok/s 50 tok/s 31 tok/s | ~1.8s ~1.8s ~1.9s | |
#208 | Qwen3.5-0.8B | 0.8B | INT4 Q8 FP16 | 186 tok/s 149 tok/s 104 tok/s | ~500 ms ~501 ms ~504 ms | |
#209 | Gemma 2 9B | 9B | INT4 | 44 tok/s | ~5.0s | |
#210 | Qwen3-14B | 14B | INT4 | 29 tok/s | ~8.4s | |
#211 | Gemma 3 4B | 4B | INT4 Q8 FP16 | 81 tok/s 53 tok/s 28 tok/s | ~2.3s ~2.3s ~2.5s | |
#212 | Qwen3-8B | 8B | INT4 Q8 | 48 tok/s 26 tok/s | ~4.5s ~5.3s | |
#214 | DeepSeek-R1 7B | 7B | INT4 Q8 | 55 tok/s 32 tok/s | ~3.9s ~4.3s | |
#215 | Qwen3-1.7B | 1.7B | INT4 Q8 FP16 | 134 tok/s 97 tok/s 60 tok/s | ~1.0s ~1.1s ~1.1s | |
#216 | Gemma 3 270M | 0.27B | INT4 Q8 FP16 | 218 tok/s 200 tok/s 167 tok/s | ~168 ms ~169 ms ~170 ms | |
#217 | Llama 3.1 8B | 8B | INT4 Q8 | 48 tok/s 26 tok/s | ~4.5s ~5.3s | |
#218 | Qwen3-0.6B | 0.6B | INT4 Q8 FP16 | 205 tok/s 170 tok/s 124 tok/s | ~371 ms ~372 ms ~374 ms | |
#220 | Llama 3 8B | 8B | INT4 Q8 | 48 tok/s 26 tok/s | ~4.5s ~5.3s | |
#222 | Phi-4-Mini | 3.8B | INT4 Q8 FP16 | 84 tok/s 55 tok/s 30 tok/s | ~2.2s ~2.2s ~2.4s | |
#224 | Qwen2.5-7B | 7B | INT4 Q8 | 54 tok/s 31 tok/s | ~3.9s ~4.3s | |
#225 | Ministral-8B-2410 | 8B | INT4 Q8 | 48 tok/s 26 tok/s | ~4.5s ~5.3s | |
#226 | DeepSeek-R1 1.5B | 1.5B | INT4 Q8 FP16 | 142 tok/s 105 tok/s 67 tok/s | ~927 ms ~930 ms ~935 ms | |
#227 | Llama 3.2 3B | 3B | INT4 Q8 FP16 | 98 tok/s 66 tok/s 38 tok/s | ~1.8s ~1.8s ~1.8s | |
#229 | Phi-3-mini | 3.8B | INT4 Q8 FP16 | 84 tok/s 55 tok/s 30 tok/s | ~2.2s ~2.2s ~2.4s | |
#230 | Gemma 3n E2B IT | 6B (2B active) | INT4 Q8 | 65 tok/s 52 tok/s | ~1.3s ~1.4s | |
#231 | Mistral-7B-Instruct-v0.1 | 7.3B | INT4 Q8 | 52 tok/s 30 tok/s | ~4.1s ~4.5s | |
#232 | Llama 3.2 1B | 1B | INT4 Q8 FP16 | 169 tok/s 133 tok/s 89 tok/s | ~634 ms ~636 ms ~640 ms | |
#233 | Gemma 2 2B | 2B | INT4 Q8 FP16 | 122 tok/s 87 tok/s 53 tok/s | ~1.2s ~1.2s ~1.2s | |
#234 | Phi-3-medium | 14B | INT4 | 29 tok/s | ~8.4s | |
#235 | OLMo 3 7B Instruct | 7B | INT4 Q8 | 38 tok/s 24 tok/s | ~3.9s ~4.6s | |
#237 | Qwen2-7B | 7B | INT4 Q8 | 54 tok/s 31 tok/s | ~3.9s ~4.3s | |
#238 | Phi-3-small | 7B | INT4 Q8 | 54 tok/s 31 tok/s | ~3.9s ~4.3s | |
#239 | Gemma 3 1B | 1B | INT4 Q8 FP16 | 169 tok/s 133 tok/s 89 tok/s | ~634 ms ~636 ms ~640 ms | |
#240 | Mistral-7B-Instruct-v0.2 | 7.3B | INT4 Q8 | 52 tok/s 30 tok/s | ~4.1s ~4.5s | |
#242 | Gemma 1 7B | 7B | INT4 Q8 | 38 tok/s 24 tok/s | ~3.9s ~4.6s | |
#243 | Gemma 1 2B | 2B | INT4 Q8 FP16 | 131 tok/s 91 tok/s 54 tok/s | ~1.2s ~1.2s ~1.2s | |
#244 | ChatGLM3-6B | 6B | INT4 Q8 | 43 tok/s 29 tok/s | ~3.4s ~3.7s | |
#245 | ChatGLM2-6B | 6B | INT4 Q8 | 43 tok/s 29 tok/s | ~3.4s ~3.7s | |
#246 | ChatGLM-6B | 6B | INT4 Q8 | 43 tok/s 29 tok/s | ~3.4s ~3.7s | |
未上榜模型（按推理速度排序） | ||||||
- | DINOv3 ViT-L/16 | 0.3B | INT4 Q8 FP16 | 239 tok/s 214 tok/s 173 tok/s | ~186 ms ~187 ms ~188 ms | |
- | Qwen2-0.5B | 0.5B | INT4 Q8 FP16 | 215 tok/s 183 tok/s 137 tok/s | ~308 ms ~309 ms ~310 ms | |
- | Qwen2.5-0.5B | 0.5B | INT4 Q8 FP16 | 215 tok/s 183 tok/s 137 tok/s | ~308 ms ~309 ms ~310 ms | |
- | ERNIE-4.5-0.3B | 0.3B | INT4 Q8 FP16 | 211 tok/s 193 tok/s 159 tok/s | ~187 ms ~187 ms ~188 ms | |
- | ERNIE-4.5-0.3B-Base | 0.3B | INT4 Q8 FP16 | 211 tok/s 193 tok/s 159 tok/s | ~187 ms ~187 ms ~188 ms | |
- | Liquid D1-Omni 600M | 0.6B | INT4 Q8 FP16 | 205 tok/s 170 tok/s 124 tok/s | ~371 ms ~372 ms ~374 ms | |
- | PixArt-α XL 2 1024-MS | 0.61B | INT4 Q8 FP16 | 204 tok/s 169 tok/s 123 tok/s | ~376 ms ~377 ms ~380 ms | |
- | PixArt-Σ XL 2 1024-MS | 0.61B | INT4 Q8 FP16 | 204 tok/s 169 tok/s 123 tok/s | ~376 ms ~377 ms ~380 ms | |
- | Nemotron 3.5 Parse | 0.63B | INT4 Q8 FP16 | 201 tok/s 167 tok/s 120 tok/s | ~391 ms ~392 ms ~394 ms | |
- | SD 2.1 Turbo | 0.86B | INT4 Q8 FP16 | 181 tok/s 144 tok/s 99 tok/s | ~539 ms ~540 ms ~544 ms | |
- | Stable Diffusion 1.4 | 0.86B | INT4 Q8 FP16 | 181 tok/s 144 tok/s 99 tok/s | ~539 ms ~540 ms ~544 ms | |
- | Stable Diffusion 2.1 | 0.86B | INT4 Q8 FP16 | 181 tok/s 144 tok/s 99 tok/s | ~539 ms ~540 ms ~544 ms | |
- | Falcon-1B | 1B | INT4 Q8 FP16 | 179 tok/s 138 tok/s 91 tok/s | ~634 ms ~636 ms ~639 ms | |
- | Falcon3-1B | 1B | INT4 Q8 FP16 | 169 tok/s 133 tok/s 89 tok/s | ~634 ms ~636 ms ~640 ms | |
- | MiniCPM-V 4.7 1B | 1B | INT4 Q8 FP16 | 169 tok/s 133 tok/s 89 tok/s | ~634 ms ~636 ms ~640 ms | |
- | Kandinsky 2.2 Decoder | 1.22B | INT4 Q8 FP16 | 156 tok/s 119 tok/s 78 tok/s | ~764 ms ~766 ms ~770 ms | |
- | Wan 2.1 T2V 1.3B | 1.3B | INT4 Q8 FP16 | 152 tok/s 115 tok/s 74 tok/s | ~808 ms ~810 ms ~815 ms | |
- | Qwen2-1.5B | 1.5B | INT4 Q8 FP16 | 142 tok/s 105 tok/s 67 tok/s | ~927 ms ~930 ms ~935 ms | |
- | Qwen2.5-1.5B | 1.5B | INT4 Q8 FP16 | 142 tok/s 105 tok/s 67 tok/s | ~927 ms ~930 ms ~935 ms | |
- | SVD Image-to-Video XT | 1.5B | INT4 Q8 FP16 | 142 tok/s 105 tok/s 67 tok/s | ~927 ms ~930 ms ~935 ms | |
- | SVD Image-to-Video XT 1.1 | 1.5B | INT4 Q8 FP16 | 142 tok/s 105 tok/s 67 tok/s | ~927 ms ~930 ms ~935 ms | |
- | LFM3 1.7B | 1.7B | INT4 Q8 FP16 | 134 tok/s 97 tok/s 60 tok/s | ~1.0s ~1.1s ~1.1s | |
- | LFM3 1.7B Base | 1.7B | INT4 Q8 FP16 | 134 tok/s 97 tok/s 60 tok/s | ~1.0s ~1.1s ~1.1s | |
- | Spark X2.5 1.7B | 1.7B | INT4 Q8 FP16 | 134 tok/s 97 tok/s 60 tok/s | ~1.0s ~1.1s ~1.1s | |
- | Hy-MT2 1.8B | 1.8B | INT4 Q8 FP16 | 130 tok/s 94 tok/s 58 tok/s | ~1.1s ~1.1s ~1.1s | |
- | Moondream 3.1 | 9B (2B active) | INT4 Q8 | 123 tok/s 79 tok/s | ~1.4s ~1.6s | |
- | CogVideoX 2B | 2B | INT4 Q8 FP16 | 122 tok/s 87 tok/s 53 tok/s | ~1.2s ~1.2s ~1.2s | |
- | LFM3 Audio 2B | 2B | INT4 Q8 FP16 | 122 tok/s 87 tok/s 53 tok/s | ~1.2s ~1.2s ~1.2s | |
- | LFM3 Omni 2B | 2B | INT4 Q8 FP16 | 122 tok/s 87 tok/s 53 tok/s | ~1.2s ~1.2s ~1.2s | |
- | LFM3 VL 2B | 2B | INT4 Q8 FP16 | 122 tok/s 87 tok/s 53 tok/s | ~1.2s ~1.2s ~1.2s | |
- | Liquid D1 2B | 2B | INT4 Q8 FP16 | 122 tok/s 87 tok/s 53 tok/s | ~1.2s ~1.2s ~1.2s | |
- | LTX-Video 0.9.1 | 2B | INT4 Q8 FP16 | 122 tok/s 87 tok/s 53 tok/s | ~1.2s ~1.2s ~1.2s | |
- | Stable Diffusion 3 Medium | 2B | INT4 Q8 FP16 | 122 tok/s 87 tok/s 53 tok/s | ~1.2s ~1.2s ~1.2s | |
- | Kandinsky 3.0 | 2.1B | INT4 Q8 FP16 | 119 tok/s 84 tok/s 51 tok/s | ~1.3s ~1.3s ~1.3s | |
- | LTX-2 | 2.2B | INT4 Q8 FP16 | 117 tok/s 82 tok/s 49 tok/s | ~1.3s ~1.3s ~1.3s | |
- | SDXL 1.0 Refiner | 2.31B | INT4 Q8 FP16 | 114 tok/s 79 tok/s 47 tok/s | ~1.4s ~1.4s ~1.4s | |
- | Stable Diffusion 3.5 Medium | 2.5B | INT4 Q8 FP16 | 109 tok/s 75 tok/s 44 tok/s | ~1.5s ~1.5s ~1.5s | |
- | SDXL 1.0 Base | 2.56B | INT4 Q8 FP16 | 107 tok/s 74 tok/s 44 tok/s | ~1.5s ~1.5s ~1.5s | |
- | SDXL Turbo | 2.56B | INT4 Q8 FP16 | 107 tok/s 74 tok/s 44 tok/s | ~1.5s ~1.5s ~1.5s | |
- | Falcon-3B | 3B | INT4 Q8 FP16 | 103 tok/s 68 tok/s 39 tok/s | ~1.8s ~1.8s ~1.8s | |
- | CroissantLLM Base | 1.3B | INT4 Q8 FP16 | 102 tok/s 85 tok/s 61 tok/s | ~811 ms ~813 ms ~818 ms | |
- | Phi-1 | 1.3B | INT4 Q8 FP16 | 102 tok/s 85 tok/s 61 tok/s | ~811 ms ~813 ms ~818 ms | |
- | Phi-1.5 | 1.3B | INT4 Q8 FP16 | 102 tok/s 85 tok/s 61 tok/s | ~811 ms ~813 ms ~818 ms | |
- | DeepSeek-R1 3B | 3B | INT4 Q8 FP16 | 99 tok/s 66 tok/s 38 tok/s | ~1.8s ~1.8s ~1.8s | |
- | Mellum 2.1 12B A2.5B Thinking | 12B (2.5B active) | INT4 | 99 tok/s | ~1.7s | |
- | Falcon3-3B | 3B | INT4 Q8 FP16 | 98 tok/s 66 tok/s 38 tok/s | ~1.8s ~1.8s ~1.8s | |
- | LFM2-VL 3B | 3B | INT4 Q8 FP16 | 98 tok/s 66 tok/s 38 tok/s | ~1.8s ~1.8s ~1.8s | |
- | Liquid D1 3B | 3B | INT4 Q8 FP16 | 98 tok/s 66 tok/s 38 tok/s | ~1.8s ~1.8s ~1.8s | |
- | Liquid D1-VL 3B | 3B | INT4 Q8 FP16 | 98 tok/s 66 tok/s 38 tok/s | ~1.8s ~1.8s ~1.8s | |
- | Liquid LFM2-VL 3B D1 | 3B | INT4 Q8 FP16 | 98 tok/s 66 tok/s 38 tok/s | ~1.8s ~1.8s ~1.8s | |
- | Qwen2.5-3B | 3B | INT4 Q8 FP16 | 98 tok/s 66 tok/s 38 tok/s | ~1.8s ~1.8s ~1.8s | |
- | FLUX.2 Klein 4B | 4B | INT4 Q8 FP16 | 81 tok/s 53 tok/s 28 tok/s | ~2.3s ~2.3s ~2.5s | |
- | FrogNano 4B | 4B | INT4 Q8 FP16 | 81 tok/s 53 tok/s 28 tok/s | ~2.3s ~2.3s ~2.5s | |
- | Mage-VL Cowork | 4B | INT4 Q8 FP16 | 81 tok/s 53 tok/s 28 tok/s | ~2.3s ~2.3s ~2.5s | |
- | Spark X2.5 4B | 4B | INT4 Q8 FP16 | 81 tok/s 53 tok/s 28 tok/s | ~2.3s ~2.3s ~2.5s | |
- | EmbeddingGemma 2 | 2B | INT4 Q8 FP16 | 78 tok/s 63 tok/s 43 tok/s | ~1.2s ~1.2s ~1.2s | |
- | CogVideoX 1.5 5B | 5B | INT4 Q8 | 69 tok/s 44 tok/s | ~2.8s ~2.9s | |
- | CogVideoX 5B | 5B | INT4 Q8 | 69 tok/s 44 tok/s | ~2.8s ~2.9s | |
- | Wan 2.2 TI2V 5B | 5B | INT4 Q8 | 69 tok/s 44 tok/s | ~2.8s ~2.9s | |
- | Phi-2 | 2.7B | INT4 Q8 FP16 | 68 tok/s 53 tok/s 35 tok/s | ~1.6s ~1.6s ~1.6s | |
- | MaLLaM-3B | 3B | INT4 Q8 FP16 | 65 tok/s 50 tok/s 31 tok/s | ~1.8s ~1.8s ~1.9s | |
- | SmolLM3 3B | 3B | INT4 Q8 FP16 | 65 tok/s 50 tok/s 31 tok/s | ~1.8s ~1.8s ~1.9s | |
- | Falcon-7B | 7B | INT4 Q8 | 56 tok/s 32 tok/s | ~3.9s ~4.3s | |
- | AuraFlow v0.2 | 6.8B | INT4 Q8 | 55 tok/s 32 tok/s | ~3.8s ~4.2s | |
- | AuraFlow v0.3 | 6.8B | INT4 Q8 | 55 tok/s 32 tok/s | ~3.8s ~4.2s | |
- | Falcon3-7B | 7B | INT4 Q8 | 54 tok/s 31 tok/s | ~3.9s ~4.3s | |
- | Hy-MT2 7B | 7B | INT4 Q8 | 54 tok/s 31 tok/s | ~3.9s ~4.3s | |
- | Mistral-7B-v0.1 | 7.3B | INT4 Q8 | 52 tok/s 30 tok/s | ~4.1s ~4.5s | |
- | Youtu Parsing Omni | 8B | INT4 Q8 | 50 tok/s 28 tok/s | ~4.5s ~4.9s | |
- | Stable Diffusion 3.5 Large | 8.1B | INT4 Q8 | 48 tok/s 26 tok/s | ~4.5s ~5.3s | |
- | Stable Diffusion 3.5 Large Turbo | 8.1B | INT4 Q8 | 48 tok/s 26 tok/s | ~4.5s ~5.3s | |
- | FLUX.2 Klein 9B | 9B | INT4 | 44 tok/s | ~5.0s | |
- | LensVLM 9B | 9B | INT4 | 44 tok/s | ~5.0s | |
- | MiMo V2.6 Distill Qwen 9B | 9B | INT4 | 44 tok/s | ~5.0s | |
- | ChatGLM3-6B-32K | 6B | INT4 Q8 | 43 tok/s 29 tok/s | ~3.4s ~3.7s | |
- | Yi-6B | 6B | INT4 Q8 | 43 tok/s 29 tok/s | ~3.4s ~3.7s | |
- | Falcon3-10B | 10B | INT4 | 40 tok/s | ~5.6s | |
- | Mochi 1 Preview | 10B | INT4 | 40 tok/s | ~5.6s | |
- | Falcon2-11B | 11B | INT4 | 39 tok/s | ~6.1s | |
- | Hunyuan Lite | 7B | INT4 Q8 | 38 tok/s 24 tok/s | ~3.9s ~4.6s | |
- | OLMo 3 7B Base | 7B | INT4 Q8 | 38 tok/s 24 tok/s | ~3.9s ~4.6s | |
- | SEA-LION-7B | 7.1B | INT4 Q8 | 38 tok/s 24 tok/s | ~4.0s ~4.7s | |
- | SEA-LION-7B-Instruct | 7.1B | INT4 Q8 | 38 tok/s 24 tok/s | ~4.0s ~4.7s | |
- | FLUX.1 Canny Dev | 11.9B | INT4 | 35 tok/s | ~6.6s | |
- | FLUX.1 Depth Dev | 11.9B | INT4 | 35 tok/s | ~6.6s | |
- | FLUX.1 Dev | 11.9B | INT4 | 35 tok/s | ~6.6s | |
- | FLUX.1 Fill Dev | 11.9B | INT4 | 35 tok/s | ~6.6s | |
- | FLUX.1 Schnell | 11.9B | INT4 | 35 tok/s | ~6.6s | |
- | Sahabat-AI-Llama3-8B-Instruct | 8B | INT4 | 35 tok/s | ~4.5s | |
- | Typhoon-2-8B | 8B | INT4 | 35 tok/s | ~4.5s | |
- | GLM-4-9B | 9B | INT4 | 30 tok/s | ~5.5s | |
- | GLM-4-9B-Chat | 9B | INT4 | 30 tok/s | ~5.5s | |
- | GLM-4-9B-Chat-1M | 9B | INT4 | 30 tok/s | ~5.5s | |
- | HunyuanVideo 1.5 720P T2V | 13B | INT4 | 30 tok/s | ~7.8s | |
- | Sahabat-AI-Gemma2-9B | 9.2B | INT4 | 30 tok/s | ~5.6s | |
- | Sahabat-AI-Gemma2-9B-Instruct | 9.2B | INT4 | 30 tok/s | ~5.6s | |
- | Yi-9B | 9B | INT4 | 30 tok/s | ~5.5s | |
- | Wan 2.1 I2V 14B 720P | 14B | INT4 | 29 tok/s | ~8.4s | |
- | Wan 2.1 T2V 14B | 14B | INT4 | 29 tok/s | ~8.4s | |
- | Wan 2.2 T2V A14B | 14B | INT4 | 29 tok/s | ~8.4s | |
- | FLUX.2 Dev | 15B | INT4 | 27 tok/s | ~9.0s | |
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
8B 模型
8位精度
3B 模型
16位全精度
3B 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
8B 模型
LoRA
3B 模型
全参数微调
< 1B 模型
负载与选型建议
AMD RX 6700 (10GB) 上的最佳量化精度与运行规格建议。
最佳推理区间
推荐以 FP16/Q8 精度运行 1B 至 3B 轻量级模型，或在 4-bit 量化下运行 7B/8B 模型。
显存带宽与推理速率
具备 320 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 71 tok/s，70B Q4 模型预估生成速度约 8 tok/s。