Direct answer: Xiaomi officially released and open-sourced the MiMo-V2.6 family on September 22, 2026 — three models (Pro, Flash, Pro-UltraSpeed) with API prices held at V2.5 levels. MiMo-V2.6-Pro scores 46 on the Artificial Analysis Intelligence Index, overtaking Kimi K3 and Qwen3.8 Max as the strongest open-weights model today, while still trailing closed-source leaders Claude Fable 5.1 and GPT-6 Astra. The MiMo Desktop client and paid subscriptions shipped the same day, and the MiMo ecosystem now spans MiMo Claw (Kingsoft office integration), MiMo Code and MiMo Studio.
What's in the release
- MiMo-V2.6-Pro: the flagship reasoning model — natively omni-modal, trillion-parameter, aimed at complex long-horizon projects, cybersecurity and research workloads.
- MiMo-V2.6-Flash: omni-modal, high-intelligence and low-cost, positioned for high-frequency professional use; it fully surpasses the previous MiMo-V2.5-Pro.
- MiMo-V2.6-Pro-UltraSpeed: Pro performance with up to 20× inference speed, for latency-sensitive production scenarios. Available both on the open platform API and inside MiMo Desktop.
- Pricing stays at V2.5 levels; Xiaomi claims a new domestic price-performance record — at equal intelligence, 1/20 to 1/60 the price of overseas models. API names are all-lowercase:
mimo-v2.6-pro,mimo-v2.6-flash,mimo-v2.6-pro-ultraspeed.
The training story: scaled RL in the open
MiMo-V2.6 was trained with large-scale multi-task reinforcement learning — per the official post, likely the largest RL compute ever sunk into a Chinese open-weights model. The V2.6 runs were streamed live: under 6 days of Live RL training, Flash and Pro cost roughly $0.85M and $2.62M respectively, each completing 30 steps with ~750k cumulative trajectories; average task pass rates improved 25% and 12%, and on the out-of-sample DeepSWE v1.1 benchmark scores rose ~17 points (48.8 → 65.7) and ~14 points (58.4 → 72.6). Training used 1M-token context with 3.5–3.7B tokens per step, mixing Code, General, Visual and Cyber task families across multiple harnesses.
On most Agent benchmarks, MiMo-V2.6-Pro lands close to Claude Opus 5 and GPT-5.6 Sol.
Beyond benchmarks, the official post demos "Vibe World": driving 3D open-world game generation, Blender modeling, embodied manipulation of a Franka Panda arm, and Computer-Use workflows — plus research cases (PFAS-absorbing MOF material screening; a full Lean 4 formalization of the Li–Yorke "period three implies chaos" theorem, 6,000+ lines kernel-verified) and creative output (frontend/PPT design, video with TTS narration, a ~10-instrument orchestral piece).
Open-source package
Weights and the technical report are open (Hugging Face XiaomiMiMo/mimo-v26 collection), plus MiMo-V2.6-Distill-Qwen-9B and RL research assets: 7k+ RL task environments across software engineering, vulnerability reproduction, knowledge work and web design; an end-to-end RL training framework built on verl / uni-agent / mini-swe-agent; and minimal composable mini-harnesses that decouple system prompts, tools and context management — supporting Multi-Harness Training for cross-framework generalization.
Why this matters for the harness ecosystem
The multi-harness RL recipe — training one model across many agent frameworks and open-sourcing the harnesses themselves — is a signal that model vendors are now optimizing directly for the harness layer, not just raw chat. For the projects this site tracks, two hooks stand out: MiMo has been integrated into Hermes Agent (with a limited-time free window), and Xiaomi's product line already includes MiMo Claw, which pairs flagship models with the Kingsoft office suite as a subscription. The intersection of frontier open models and agent harnesses keeps getting busier — see our glossary and compare hub.
Sources
- Xiaomi MiMo official announcement (2026-09-22, Chinese): mimo.mi.com/docs/zh-CN/news/latest/v2-6
- Model release changelog: mimo.mi.com model updates
- Open weights: huggingface.co/collections/XiaomiMiMo/mimo-v26
Direct answer: 小米于 2026 年 9 月 22 日正式发布并开源 MiMo-V2.6 系列——Pro / Flash / Pro-UltraSpeed 三款模型，API 定价与 V2.5 持平。MiMo-V2.6-Pro 在 Artificial Analysis 综合智能指数取得 46 分，超过 Kimi K3 与 Qwen3.8 Max，成为当前最强开源模型；但与最强闭源模型 Claude Fable 5.1、GPT-6 Astra 仍有差距。MiMo Desktop 桌面客户端与订阅同日上线，MiMo 产品线已覆盖 MiMo Claw、MiMo Code 与 MiMo Studio。
这次发布了什么
- MiMo-V2.6-Pro：旗舰推理模型，原生全模态、万亿参数，面向复杂项目、长程任务、网络安全与科研场景；
- MiMo-V2.6-Flash：全模态、高智能、低成本，面向专业办公高频调用，全面超越上代 MiMo-V2.5-Pro；
- MiMo-V2.6-Pro-UltraSpeed：保留 Pro 旗舰性能，提供最高 20 倍推理速度，面向强实时交互场景，API 与 MiMo Desktop 内均可使用；
- 价格沿用 V2.5：官方称刷新国产模型性价比纪录——同等智能水平下，价格为海外模型的 1/20 至 1/60。API 调用模型名全小写：
mimo-v2.6-pro、mimo-v2.6-flash、mimo-v2.6-pro-ultraspeed。
训练侧：公开跋涉的大规模 RL
官方介绍，MiMo-V2.6 的 RL 训练可能是国产开源模型中投入算力最多的之一。V2.6 训练全程 Live 分享：不到 6 天完成，Flash 与 Pro 训练成本分别约 85 万与 262 万美元，各 30 步、累计约 75 万条轨迹；训练任务平均通过率分别相对提升 25% 与 12%，样本外长程软件工程评测 DeepSWE v1.1 分别提升约 17 分（48.8 → 65.7）与约 14 分（58.4 → 72.6）。训练支持 1M 上下文、单步 3.5～3.7B Token，混合 Code、General、Visual、Cyber 多任务与多个 Harness。
在多数 Agent 基准上，MiMo-V2.6-Pro 取得比肩 Claude Opus 5 与 GPT-5.6 Sol 的效果。
官方同时展示了「Vibe World」能力：3D 开放世界游戏生成、Blender 建模、Franka Panda 机械臂具身操作与 Computer Use 办公操作；科研侧有 PFAS 吸附 MOF 材料筛选、Li–Yorke「周期三蕴含混沌」定理的 Lean 4 完整形式化（6000 余行、内核核验通过）；创作侧覆盖前端/PPT 设计、配旁白的科普视频与约十种乐器的管弦乐。
开源内容
模型权重与技术报告全部开源（Hugging Face XiaomiMiMo/mimo-v26 合集），另有 MiMo-V2.6-Distill-Qwen-9B 与 RL 研究资源：覆盖软件工程、漏洞复现、知识型工作、网页设计开发的 7k+ RL 任务环境；基于 verl / uni-agent / mini-swe-agent 的端到端 RL 训练框架；以及解耦系统提示、工具与上下文管理的极简可组合 mini-harnesses——通过 Multi-Harness Training 提升模型跨框架泛化。
对 Harness 生态意味着什么
「一个模型混多个 Harness 训练、连 Harness 本身也开源」——厂商开始直接为 Harness 层优化模型，而不只是聊天能力。对本站跟踪的项目，两个交叉点值得注意：MiMo 已接入 Hermes Agent（此前限免两周）；小米产品线中的 MiMo Claw 把旗舰模型与金山办公打包成订阅服务。前沿开源模型与 Agent Harness 的交集正越来越热闹，参见站内术语表与对比总览。
来源
- 小米 MiMo 官方公告（2026-09-22）：MiMo-V2.6：扩展强化学习规模，迈向自我提升
- 官方模型发布日志：mimo.mi.com 模型更新
- 开源权重：huggingface.co/collections/XiaomiMiMo/mimo-v26