比较并探索基于 开源模型 表现进行 Elo 积分排名的主流视觉 AI 模型，数据来源于真实用户匿名盲测。Explore Elo rankings of multimodal AI models that reason about images, charts, and OCR inputs — rated by real anonymous user votes.
每个榜单会采用官网对应的核心排序指标。通用 Arena 榜单通常是模型竞技积分；智能体榜单则展示净提升及多项任务指标。
Each leaderboard uses the official primary ranking metric. Most Arena boards use competitive scores; the Agent board uses net improvement and task-specific metrics.
主指标的完整浮点数值。对于通用 Arena 榜单可用于区分整数分相近的模型；智能体榜单以百分比和置信区间展示。
The full floating-point value of the primary metric. It helps distinguish close scores on standard boards; Agent metrics are shown as percentages with confidence intervals.
衡量结果的统计稳定程度。绿色·稳定（<5）更可信；橙色·波动（5~12）；红色·不稳（>12）说明排名还可能明显变化。
Measures statistical stability. Green·Stable (<5); Orange·Unstable (5–12); Red·Volatile (>12) means the ranking may still move significantly.
95% 置信区间反映当前估计的不确定范围，显示为 ±N。在相同计算口径下，N 越小说明估计越集中、排名越稳定。
The 95% confidence interval, shown as ±N, summarizes uncertainty in the current estimate. Under the same methodology, a smaller N indicates a more precise estimate and a more stable ranking.
该模型或厂商参与评测的总次数。数量越高，结果的统计置信度越高、越不容易受单次样本影响而波动。
Total evaluation volume for the model or lab. Higher counts mean higher statistical confidence and less sensitivity to individual samples.
Apache/Llama：完全开源可商用；MIT：开源限制极少；Proprietary：闭源商业模型。协议决定你能否把它集成到自己的产品里。
Apache/Llama: Fully open-source, commercial use allowed; MIT: Open-source, minimal restrictions; Proprietary: Closed-source commercial model. License determines if you can integrate it into your product.
| 排名Rank | 名次区间Rank Range | 模型Model | 投票数Votes | 精确分 & 区间Float Score & CI | 方差Variance | 价格 ($/1M)Price ($/1M) | 上下文Context | |
|---|---|---|---|---|---|---|---|---|
| 10 | 3 - 72 |
kimi-k2.5-thinking
月之暗面 · Modified MIT
|
1261±8.9 | 8.2K 票votes | 1261.180 [1252.3, 1270.1] | 20.58 · 不稳Volatile | $0.60 / $3 | N/A |
没有找到匹配的模型，换个关键词试试？No matching models found. Try a different keyword?
月之暗面