Benchmark Atlas

亚洲模型实验室 · arXiv 论文

Known vs Undisclosed Parameter Frontier

复现 GLM-4.5 Figure 2:已知参数模型使用连续横轴,闭源且参数未披露的模型进入断轴后的专用通道。

glmswe-benchparameterspareto
SWE-bench quality vs known and undisclosed parameters 复现 GLM-4.5 Figure 2:已知参数模型使用连续横轴,闭源且参数未披露的模型进入断轴后的专用通道。 示意数据,仅复现信息结构,不代表最新榜单结果。 SWE-bench quality vs known and undisclosed parameters GLM-4.5 Technical Report · Figure 2 / PDF p.3 DEMO DATA 0250500750100040506070UNDISCLOSEDGLM-4.5-AirGLM-4.5MiniMax-M1DeepSeek-R1Qwen3Kimi K2ClaudeGeminiGPT-4.1known model parameters (B) →SWE-bench Verified ↑
Illustrative demo data · renderer: unknownParameterFrontier

Information grammar

它如何组织信息

连续参数 Pareto 区 + 断轴后的 undisclosed 垂直通道 + 前沿楔形与异形点

Visual system
VS-71 GLM mixed-domain parameter frontier
Component ID
glm-parameter-frontier-unknown-lane

Use when

适用判断

适合参数效率比较又必须保留未披露模型时;未知参数点不可被放到虚构的连续坐标。

Figure-level evidence

Figure 2 · PDF p.3

开源模型按总参数定位,专有模型在右侧 Unknown 通道比较 SWE-bench Verified。

Verified
2026-07-12
Paper source
GLM-4.5 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。

{
  "frontier": [
    [
      0,
      51
    ],
    [
      106,
      57
    ],
    [
      355,
      64
    ],
    [
      650,
      72
    ]
  ],
  "known": [
    {
      "label": "GLM-4.5-Air",
      "parameters": 106,
      "score": 57.2,
      "frontier": true,
      "dx": 10,
      "dy": -8
    },
    {
      "label": "GLM-4.5",
      "parameters": 355,
      "score": 64.2,
      "frontier": true,
      "dx": 10,
      "dy": -9
    },
    {
      "label": "MiniMax-M1",
      "parameters": 456,
      "score": 55.7,
      "frontier": false,
      "dx": 10,
      "dy": 16
    },
    {
      "label": "DeepSeek-R1",
      "parameters": 671,
      "score": 57.3,
      "frontier": false,
      "dx": 10,
      "dy": -8
    },
    {
      "label": "Qwen3",
      "parameters": 235,
      "score": 35.6,
      "frontier": false,
      "dx": 10,
      "dy": -8
    },
    {
      "label": "Kimi K2",
      "parameters": 1043,
      "score": 64.9,
      "frontier": false,
      "dx": -10,
      "dy": -8
    }
  ],
  "unknown": [
    {
      "label": "Claude",
      "score": 69.8
    },
    {
      "label": "Gemini",
      "score": 62.4
    },
    {
      "label": "GPT-4.1",
      "score": 54.8
    }
  ]
}

Related components

同一图表家族的其他语法