大模型
www.aibase.com ·
小米 MiMo-V2.6-Pro 重返 Code Arena 前十，AutoEval 得分 1628
据 Arena.ai 最新评测数据，小米全模态旗舰模型 MiMo-V2.6-Pro 在 Code Arena WebDev 首次亮相即进入总榜前十，在 MIT 许可的开源权重模型中位列前三。其 AutoEval 得分为 1628，与 Claude Fable5 (High) 持平，略高于 Hy4-preview 的 1624，较上代 MiMo-V2.5-Pro 的 1475 提升 153 分；在开源权重模型中落后 Qwen3.8Flash Next 约 7 分、Kimi K3Max 46 分。该结果目前来自基于人类偏好数据训练的奖励模型自动投票，尚待更多真实人类投票数据验证。
AI推荐理由 提供了具体可核验的评测数字：MiMo-V2.6-Pro 在 Code Arena WebDev 的 AutoEval 得分 1628，与 Claude Fable5 (High) 持平、略高于 Hy4-preview 的 1624，较上代 MiMo-V2.5-Pro 的 1475 提升 153 分，并落后 Qwen3.8Flash Next 约 7 分、Kimi K3Max 46 分。这些横向对比对关注开源模型竞争格局的开发者有参考价值。局限在于：来源为 AIbase 转述 Arena.ai 评测，非小米或 Arena.ai 官方发布；文中明确该成绩来自基于人类偏好数据训练的奖励模型自动投票，最终排名仍需真实人类投票验证，因此不能视为已确认的稳定排名；正文未给出评测方法细节、模型参数或开源权重链接。
62
正文
Open-Source Large Model Landscape Reimagined! Xiaomi MiMo-V2.6-Pro Makes a Strong Return to the Top Ten in Code Arena, Performance Approaching the World's Leading Tier
AIbase基地
Published inAI News · 3 min read · Sep 28, 2026
4
The AI open-source ecosystem has welcomed another strong contender. According to the latest evaluation data released by Arena.ai, Xiaomi's new all-modal flagship model, MiMo-V2.6-Pro, performed outstandingly in the Code Arena: WebDev (Web Development Test Field). Upon its debut, it successfully returned to the top ten of the overall ranking and ranked among the top three in open-source weight models using the MIT license.
Regarding specific scoring performance, MiMo-V2.6-Pro achieved a high score of 1628 in early automated evaluation (AutoEval). This result is not only on par with Claude Fable5 (High), but also slightly surpasses Hy4-preview, which scored 1624. Compared to the previous generation product MiMo-V2.5-Pro, which previously scored 1475, this represents a significant leap of 153 points in a single iteration.
In the horizontal comparison of open-source weight models, the model also showed excellent performance. Currently, its score lags behind the second-ranked Qwen3.8Flash Next by about 7 points, and is only 46 points behind the top-ranked Kimi K3Max.
It should be noted that this result is currently derived from automated voting based on a reward model trained on human preference data. As more real human voting data continues to accumulate and converge, the final ranking and performance of the model will continue to be verified. The strong debut of MiMo-V2.6-Pro once again demonstrates Xiaomi's continuous efforts and technical breakthroughs in the fields of reinforcement learning and all-modal large model development.
This article is from AIbase Daily
Scan to view
Welcome to the [AI Daily] column! This is your daily guide to exploring the world of artificial intelligence. Every day, we present you with hot topics in the AI field, focusing on developers, helping you understand technical trends, and learning about innovative AI product applications.
—— Created by the AIbase Daily Team
© Copyright AIbase Base 2024, Click to View Source -