通用计算单位 · 用计元量化不同AI模型的真实推理成本?
鼠标悬停模型名查看信息 · 点击?查看术语说明 · 计元 = 1 GFLOPs ?
| # | 模型 | 架构 | 活跃参数? | 词表? | 综合效率? chars/tok |
中文? chars/tok |
英文? chars/tok |
代码? chars/tok |
输入? 4K(计元) |
输出? 4K(计元) |
输入? 131K(计元) |
输出? 131K(计元) |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| #1 | DeepSeek-V3 | MoE | 37.0B | 128,815 | 3.74 | 2.03 | 6.68 | 3.67 | 88.8 | 222.0 | 159.8 | 399.6 |
| #2 | DeepSeek-R1 | MoE | 37.0B | 128,815 | 3.74 | 2.03 | 6.68 | 3.67 | 88.8 | 222.0 | 159.8 | 399.6 |
| #3 | InternLM2.5-7B | Dense | 7.0B | 92,544 | 3.71 | 1.80 | 6.68 | 4.08 | 28.0 | 70.0 | 50.4 | 126.0 |
| #4 | Qwen2.5-72B | Dense | 72.0B | 151,665 | 3.71 | 1.88 | 6.68 | 4.16 | 144.0 | 360.0 | 259.2 | 648.0 |
| #5 | Qwen2.5-7B | Dense | 7.2B | 151,665 | 3.71 | 1.88 | 6.68 | 4.16 | 14.4 | 36.0 | 25.9 | 64.8 |
| #6 | Qwen2.5-0.5B | Dense | 0.5B | 151,665 | 3.71 | 1.88 | 6.68 | 4.16 | 2.0 | 5.0 | 3.6 | 9.0 |
| #7 | Baichuan2-7B | Dense | 7.0B | 125,696 | 3.48 | 2.00 | 6.60 | 2.77 | 28.0 | 70.0 | 50.4 | 126.0 |
| #8 | Yi-1.5-6B | Dense | 6.0B | 63,992 | 3.48 | 1.86 | 6.60 | 2.98 | 24.0 | 60.0 | 43.2 | 108.0 |
| #9 | MiniCPM3-4B | Dense | 3.6B | 73,448 | 3.38 | 1.80 | 6.46 | 2.89 | 14.4 | 36.0 | 25.9 | 64.8 |
| #10 | Phi-4 | Dense | 14.0B | 100,352 | 3.12 | 0.93 | 6.68 | 4.18 | 56.0 | 140.0 | 100.8 | 252.0 |
| #11 | Mistral-7B | Dense | 7.0B | 32,768 | 2.74 | 0.95 | 5.99 | 2.98 | 28.0 | 70.0 | 50.4 | 126.0 |