亚洲模型实验室 · arXiv 论文
Isoflop Equal-Quality Gaps
复现 MiniMax-01 Figure 4：在多个 benchmark 中，以同一性能水平反查 MoE 与 dense 所需算力差。
isoflopGapFacetsInformation grammar
它如何组织信息
benchmark 小多图双拟合曲线 + 等性能水平桥 + 两端算力垂线与倍率
- Visual system
- VS-65 MiniMax isoflop gap facets
- Component ID
minimax-isoflop-gaps
Use when
适用判断
适合回答“达到同一质量要多少 compute”；不同训练 token 或激活参数设置不可混在同一桥线。
Figure-level evidence
Figure 4 · PDF p.5
五个 benchmark facet 比较 2B-active MoE 与 7B dense 的等性能算力差。
- Verified
- 2026-07-12
- Paper source
- MiniMax-01 Technical Report
Structured data
可替换的数据模型
示意数据，仅复现信息结构，不代表最新榜单结果。
{
"facets": [
{
"label": "HellaSwag",
"min": 54,
"max": 74,
"target": 70,
"moeMatch": 3.4,
"denseMatch": 10.2,
"moe": [
[
0.5,
58
],
[
1.2,
63
],
[
2.1,
67
],
[
3.4,
70
],
[
4.2,
72
]
],
"dense": [
[
0.5,
55
],
[
2,
60
],
[
5,
66
],
[
9,
69
],
[
13,
72
]
]
},
{
"label": "WinoGrande",
"min": 56,
"max": 69,
"target": 65,
"moeMatch": 3.1,
"denseMatch": 9.6,
"moe": [
[
0.5,
58
],
[
1.3,
61
],
[
2.2,
63
],
[
3.1,
65
],
[
4.1,
66
]
],
"dense": [
[
0.5,
57
],
[
2.5,
60
],
[
5.5,
63
],
[
9.6,
65
],
[
13,
67
]
]
},
{
"label": "Natural Questions",
"min": 5,
"max": 16,
"target": 13,
"moeMatch": 3.7,
"denseMatch": 11.6,
"moe": [
[
0.5,
7
],
[
1.3,
9
],
[
2.3,
11
],
[
3.7,
13
],
[
4.4,
15
]
],
"dense": [
[
0.5,
6
],
[
2.8,
8.5
],
[
6,
10
],
[
9,
11
],
[
11.6,
13
],
[
13.5,
14.5
]
]
},
{
"label": "PIQA",
"min": 70,
"max": 80,
"target": 77,
"moeMatch": 2.8,
"denseMatch": 11.8,
"moe": [
[
0.5,
73
],
[
1.2,
75
],
[
2,
76
],
[
2.8,
77
],
[
3.8,
79
]
],
"dense": [
[
0.5,
71.5
],
[
3,
74
],
[
6,
75
],
[
9,
76
],
[
11.8,
77
],
[
13.5,
77.4
]
]
}
]
}Related components