实时更新 · 官方数据Live Updates · Official Data

LM Arena 大模型全球实力榜LM Arena Global AI Leaderboard困难模式🌶️Hard Mode

Name: Stability AI 困难模式 · 大语言模型排行榜 - 最佳文本与对话 AI 模型对比 2026-04
Creator: LM Arena
License: https://creativecommons.org/licenses/by/4.0/
Keywords: Stability AI,困难模式,AI 大语言模型排行榜,AI 模型排行

查看各种 AI 大语言模型在数学推理、代码生成、创意写作及其他开放式文本任务中的综合 Elo 积分排名，数据来源于真实用户的匿名盲测投票。

2026-04-22

▾

1

上榜模型Models

1

参赛厂商Providers

575

总投票数Total Votes

🥇 #1 排名Rank

stablelm-tuned-alpha-7b

Stability AI

941

±25 · 575 次投票votes

📖 指标说明 — 读懂这张榜单📖 Metrics Guide — Understanding the Leaderboard

🎯

Arena 得分（Elo Rating）Arena Score (Elo Rating)

基于 Bradley-Terry 统计模型计算的大模型竞技积分。数值越高代表在人类盲测对决中综合胜率越高，反映模型间相对强弱。

Competitive score based on the Bradley-Terry statistical model. Higher score = better overall win rate in human blind tests, reflecting relative model strength.

🔢

精确分（Float Rating）Float Rating

Elo 积分的完整浮点数值（精确到小数3位）。当两个模型整数分相同时，精确分才真正决定排名先后。下方 [lower, upper] 为置信区间范围。

Full floating-point Elo score (3 decimal places). When two models share the same integer score, Float Rating determines the final ranking. [lower, upper] shows the confidence interval.

📊

方差（Variance）Variance

衡量 Elo 积分的统计稳定程度。绿色·稳定（<5）票多可信；橙色·波动（5~12）；红色·不稳（>12）新模型票少，排名还会剧烈变化。

Measures statistical stability of the Elo score. Green·Stable (<5); Orange·Unstable (5–12); Red·Volatile (>12) — new models with fewer votes may fluctuate significantly.

↔️

置信区间（95% CI = ±N）95% Confidence Interval (CI = ±N)

真实实力有 95% 概率落在 [lower, upper] 区间内，显示为 ±N。N 越小排名越可信，票数越多 N 越小。

The model's true strength falls within [lower, upper] with 95% probability, shown as ±N. Smaller N = more reliable ranking; more votes → smaller N.

🗳️

投票数（Vote Count）Vote Count

该模型参与盲测对决的总次数。票数越高，Elo 分的统计置信度越高、越不容易受单次投票影响而波动。

Total number of head-to-head comparisons this model participated in. Higher vote count = higher statistical confidence in the Elo score.

📜

开源协议（License）License

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.