MiMo V2.6（小米 MiMo） MiMo V2.6 (Xiaomi MiMo)

来源Sources

为什么推荐Why we recommend it 收录理由：2026-09-22 HN 首页出现的小米 MiMo 新系列旗舰，官方页 + HF 开源卡（MIT）+ AA 第三方榜同日可核，适合 Models 看板头条。核实边界：架构参数、模态、1M 上下文、RL 主题与系列定位一律为官方自述（取自 mimo.xiaomi.com/mimo-v2-6 meta 与 HF 卡 XiaomiMiMo/MiMo-V2.6-Pro-RL）；Intelligence Index 46 及 AA 页速度/定价为 Artificial Analysis 第三方榜数据，非本库复跑；HF 卡内厂商对照表分数不在本卡复述。本库未实测 API / Studio。整理质量 4 星（来源齐、边界清），不代表基准分。 Why: Xiaomi MiMo’s new-series flagship landing on the HN front page on 2026-09-22, with a verifiable official page, MIT HF card, and same-day AA third-party coverage — a strong Models-board headline. Boundary: architecture, modalities, 1M context, RL theme, and series positioning are vendor claims from mimo.xiaomi.com/mimo-v2-6 meta and the HF card XiaomiMiMo/MiMo-V2.6-Pro-RL; Intelligence Index 46 and AA speed/pricing are Artificial Analysis board figures, not re-run here; individual vendor-table scores from the HF card are not restated on this card. Official page, HF Pro card, Flash/Distill sibling repos, AA, API platform, and HN items 49792730 / 49796660 all returned 200 on 2026-09-22. API / Studio not tested here.

要点（官方称 / 社区称）Highlights (vendor / community claims)

• 官方称（HF 模型卡）：MiMo-V2.6-Pro-RL 是 MiMo-V2.6 系列旗舰，主题「Scaling Reinforcement Learning Toward Self-Improvement」——一起放大 RL 算力、环境多样性与 grader 算力，使模型通过探索与反馈扩展能力边界。 Vendor claim (HF model card): MiMo-V2.6-Pro-RL is the flagship checkpoint of the MiMo-V2.6 series, themed “Scaling Reinforcement Learning Toward Self-Improvement” — scaling RL compute, environment diversity, and grader compute together so the model keeps expanding its capability frontier through exploration and feedback.
• 官方称（HF 模型卡）：稀疏 Mixture-of-Experts，1.02T 总参数 / 42B 激活；原生全模态输入（文本、图像、视频、音频）；上下文 1M token。同系列 HF 还有 MiMo-V2.6-Flash-RL 与 MiMo-V2.6-Distill-Qwen-9B（两仓 HTTP 200）。 Vendor claim (HF model card): sparse Mixture-of-Experts with 1.02T total / 42B activated parameters; native omnimodal inputs (text, image, video, audio); 1M-token context. Series siblings on HF include MiMo-V2.6-Flash-RL and MiMo-V2.6-Distill-Qwen-9B (both repos HTTP 200).
• 官方页 meta description：「Introducing the MiMo-V2.6 series: frontier intelligence, all the modalities, built in public.」官方称权重 MIT 开源；API 平台 platform.xiaomimimo.com 当日可达。 Official-page meta description: “Introducing the MiMo-V2.6 series: frontier intelligence, all the modalities, built in public.” Vendor claims MIT open weights; the API platform at platform.xiaomimimo.com was reachable today.
• 第三方榜（Artificial Analysis，非官方、非本库）：MiMo-V2.6-Pro Intelligence Index 46；页面同时列出开源权重、MIT、1M 上下文、多模态输入（text / image / speech / video），以及约 124.5 tok/s 输出速度与 $0.435 / $0.87 每百万输入/输出 token 等 AA 自有数字——一律为 AA 榜单数据，本库未复跑。 Third-party board (Artificial Analysis — neither vendor nor this library): MiMo-V2.6-Pro Intelligence Index 46; the page also lists open weights, MIT, 1M context, multimodal inputs (text / image / speech / video), plus AA’s own ~124.5 output tok/s and $0.435 / $0.87 per 1M input/output tokens — all AA board figures, not re-run here.
• HF 模型卡另附厂商自测对照表（DeepSWE / AutomationBench / Terminal Bench 等与 Claude Opus 5、GPT-5.6 Sol、Claude Fable 5 等对照）。表内分数一律为厂商自述，本库未复跑任何一项，亦不在此复述具体分值以免与 AA 榜混读。 The HF model card also includes a vendor evaluation table (DeepSWE / AutomationBench / Terminal Bench and peers vs Claude Opus 5, GPT-5.6 Sol, Claude Fable 5, etc.). Those scores are vendor claims only — none re-run here, and individual numbers are not restated on this card to avoid mixing them with the AA board.