
IT之家 9 月 22 日消息,小米今日凌晨正式发布并开源了全新的 Xiaomi MiMo-V2.6 系列,包含 Pro 与 Flash 两个原生全模态模型。
小米称这是其探索 RSI(递归自我改进)路径的关键一步,借助规模化扩展强化学习算力,让模型在持续的探索与反馈中拓展智能边界。
得益于 RL 算力的扩展,MiMo-V2.6-Pro 在 Artificial Analysis Intelligence Index(AA 综合智能指数)v4.3.2 中取得 46 分,超过 Kimi K3(44 分)和 GLM-5.3(45 分),成为当前 AA 指数上排名最高的开源权重模型,较其前代 MiMo-V2.5-Pro 的 26 分高出 20 分。
小米官方同时说明,该模型与 Claude Fable 5.1 和 GPT-6 Astra(均为 53 分)等顶级闭源模型相比仍有差距。
在 RL 训练阶段,MiMo-V2.6 可能是当下国产开源模型中投入 RL 算力最多的模型之一。Pro 与 Flash 训练历时不到 6 天,成本分别约 262 万美元与 85 万美元,各完成 30 步,累计约 75 万条轨迹。
训练任务平均借助率分别相对提高 25% 和 12%,样本外的长程软件工程实测基准 DeepSWE v1.1 分别提高约 17 分(48.8 至 65.7)和约 14 分(58.4 至 72.6)。
本次训练从三个维度扩展 RL 算力:更大的 Batch 与更高吞吐,单次更新使用 1568 个样本,兼容 100 万上下文长度训练,单步训练 Token 达 3.5 至 3.7B;更多任务与复杂环境,构建覆盖 Code、General、Visual、Cyber 等方向的多任务训练体系;更大的 Grader 算力,借助 Group 内相比较较为长程 RL 任务给出更精准的奖励信号。
能力扩展上,MiMo-V2.6 融合了 3D 空间推理、多模态感知与计算机操作能力。在游戏开发中,使用者输入图片、视频或文字后,模型可将需要拆解为多个任务,由多智能体协作完成 3D 情形搭建、交互逻辑编写与视觉验证。在具身仿真环境中,MiMo-V2.6 可直接以多视角相机画面为输入,借助视觉反馈闭环控制 Franka Panda 机械臂,完成物体抓取、颜色匹配与精准放置。IT之家注:全模态模型指可同时处理文本、图像、视频和音频等多种输入形式的模型。
科研维度,MiMo-V2.6-Pro 协助研究人员完成了全氟和多氟烷基物质(PFAS)吸附用金属有机框架(MOF)材料的外观思路筛选,并在 Lean 4 中完成了 Li-Yorke 经典论文《周期三蕴含混沌》原始主定理的完整形式化,形成 6000 余行 Lean 源码,完整证明借助 Lean 内核核验。模型未接受过针对 Lean 的专项后训练。
MiMo-V2.6 系列沿用 V2.5 系列的 API 定价:Flash 为每百万词元输入 1 元、每百万输出 2 元;Pro 为 3 元和 6 元,并给出 99% 缓存折扣。在同等智能水平下,售价仅为海外模型的 1/20 至 1/60。
伴随新模型发布,MiMo Desktop 桌面客户端(兼容 Windows 与 Mac)及会员订阅思路同步上线,订阅会员即可使用 MiMo-V2.6-Pro 和 Flash 模型,也可规格自己的 API Key 使用客户端。
MiMo Desktop 同步上线 MiMo-V2.6-Pro 的 UltraSpeed 超高速模式,给出最高 20 倍的推理速度。
MiMo-V2.6 系列已上线 Xiaomi MiMo 开放平台。原邀测活动将在 1 周后结束,已取得邀测资格的使用者切换模型名称后方可继续使用。
同时,小米全面开源 MiMo-V2.6-Pro、Flash 模型权重与工艺报告,同步开放 MiMo-V2.6-Distill-Qwen-9B 及配套 RL 研究资源,分享经过验证的训练实践,以下是详细开源内容:
本文来源:https://www.ithome.com/1/005/496.htm
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。
Home > Intelligent Age > Artificial Intelligence
Thanks to IT Home netizens Langxi Qingnian, Domado, DonaSarkar_, Xingluo Wuying, Shimmianshiming, Qinglang de Chenfeng, Shunshi Erwei, Xinghe zhi Meng, Mianbao Jia Paomian, Zixin de Meimao, DayDayUp, and StarDevOps for their tip submissions!
On September 22, Xiaomi officially released and open-sourced the new Xiaomi MiMo-V2.6 series, which includes the Pro and Flash versions of the native full-modal model.
Xiaomi stated that this is a key step in their exploration of the RSI (Recursive Self-Improvement) path, utilizing large-scale expanded reinforcement learning computing power to allow the model to expand the boundaries of intelligence through continuous exploration and feedback.
Thanks to the expansion of RL computing power, MiMo-V2.6-Pro achieved a score of 46 on the Artificial Analysis Intelligence Index (AA Comprehensive Intelligence Index) v4.3.2, surpassing Kimi K3 (44 points) and GLM-5.3 (45 points), becoming the highest-ranking open-source weighted model on the AA index currently, 20 points higher than its predecessor MiMo-V2.5-Pro.
The official Xiaomi note also indicates that the model still has a gap compared to top closed-source models like Claude Fable 5.1 and GPT-6 Astra (both at 53 points).
During the RL training phase, MiMo-V2.6 is likely one of the most invested models in RL computing power among domestic open-source models. The Pro and Flash training lasted less than 6 days, costing approximately $2.62 million and $850,000 respectively, completing 30 steps each, totaling about 750,000 trajectories.
The average utilization of training tasks increased by 25% and 12% respectively, and the out-of-sample long-range software engineering benchmark DeepSWE v1.1 improved by approximately 17 points (from 48.8 to 65.7) and 14 points (from 58.4 to 72.6) respectively.
This training expanded RL computing power in three dimensions: larger batch and higher throughput, using 1568 samples per update, compatible with 1 million context length training, with single-step training tokens reaching 3.5 to 3.7 billion; more tasks and complex environments, constructing a multi-task training system covering Code, General, Visual, and Cyber; and greater grader computing power, providing more accurate reward signals for longer RL tasks within groups.
In terms of capability expansion, MiMo-V2.6 integrates 3D space reasoning, multi-modal perception, and computer operation capabilities. In game development, after users input images, videos, or text, the model can decompose the requirements into multiple tasks, which are then collaboratively completed by multi-agents for 3D scene building, interaction logic writing, and visual verification. In embodied simulation environments, MiMo-V2.6 can directly use multi-view camera images as input to control the Franka Panda robotic arm through visual feedback loops, completing object grasping, color matching, and precise placement. IT之家 note: A full-modal model refers to one that can process various input forms such as text, images, videos, and audio simultaneously.
In the field of research, MiMo-V2.6-Pro assisted scientists in screening the appearance and structure of metal-organic framework (MOF) materials for the adsorption of per- and polyfluoroalkyl substances (PFAS), and completed the full formalization of the classic Li-Yorke paper "Period Three Implies Chaos" in Lean 4, generating over 6000 lines of Lean source code, with the full proof verified by the Lean kernel. The model had no specific post-training on Lean.
The MiMo-V2.6 series retains the API pricing of the V2.5 series: Flash is 1 yuan per million input tokens and 2 yuan per million output tokens; Pro is 3 and 6 yuan respectively, with a 99% cache discount. At the same level of intelligence, the price is only 1/20 to 1/60 of overseas models.
With the release of the new model, the MiMo Desktop application (compatible with Windows and Mac) and member subscription service are also launched, allowing subscribers to use the MiMo-V2.6-Pro and Flash models or use their own API keys with the client.
MiMo Desktop simultaneously launches the UltraSpeed mode for MiMo-V2.6-Pro, providing up to 20 times the inference speed.
The MiMo-V2.6 series is now available on the Xiaomi MiMo Open Platform. The original invitation-only testing event will end in one week, and testers who have already gained access can continue using by switching the model name.
Meanwhile, Xiaomi has fully open-sourced the weights and process reports for MiMo-V2.6-Pro and Flash, and shared the MiMo-V2.6-Distill-Qwen-9B and related RL research resources, presenting validated training practices, which are detailed below:








