为了推广自家的大模型，之前搞token大赠送，然后吸引开发者主动申请，泄露自己的邮箱，然后现在小米就可以光明正大的给这些邮箱发促销信息了。
| 尊敬的开发者，感谢您一直以来对 Xiaomi MiMo 的支持！Xiaomi MiMo 团队持续优化推理性能，现将阶段性的降本成果回馈给开发者。MiMo-V2.5 全系大幅调价，最高降幅 99%新价格将于北京时间 2026.05.27 00:00 正式生效，不区分输入长度区间。V2 系列模型价格不变且即将下线，建议尽快切换新版模型。MiMo-V2.5-Pro（百万 tokens）输入（命中缓存）：¥0.025输入（未命中缓存）：¥3输出：¥6MiMo-V2.5（百万 tokens）输入（命中缓存）：¥0.02输入（未命中缓存）：¥1输出：¥2MiMo-V2.5-TTS 系列 继续限时免费查看完整定价 →Token Plan 加量不加价Credits 加量不加价：V2.5 系列模型用量可提升 5–8 倍；对 cache、输入、输出整体比例均有计量优化，整体更清晰。Credits 用量再重置：所有仍在有效期的 Token Plan（包括参与百万亿 Token 创造者激励计划获得的 Token Plan，涵盖 Apache 软件基金会专属福利），无论当前套餐的用量如何，其已消耗的 Credits 额度将被完全重置，有效期不变。One More Thing：针对 Token Plan 已过期的历史付费用户，我们也同样准备了惊喜好礼，将在未来一周宣布，请保持关注。了解 Token Plan →「MiMo Orbit：百万亿 Token 创造者激励计划」圆满收官激励计划自 2026.04.28 上线以来，受到全球用户的热情关注和广泛参与，截至北京时间 2026.05.26 16:08，100T Tokens 已全部发放完毕，活动提前收官、圆满结束，感谢广大开发者的踊跃支持！查看详情 →推理技术优化说明本次价格调整背后，离不开小米技术团队在推理系统上的持续优化。我们基于 SGLang HiCache 完整支持 SWA（Sliding Window Attention），将 KV Cache 在 GPU 显存、CPU 内存、SSD 等多级存储之间的数据搬运量降低至优化前的近 1/7，并将可缓存 token 数量提升至优化前的近 5 倍，显著提升了缓存命中率和推理效率。同时，我们通过优化专家并行方案、输入长度分桶策略等，进一步提升了集群输入吞吐能力，从而在保障服务质量的前提下持续降低单位 token 服务成本。——Xiaomi MiMo API 开放平台团队 |
| Dear Developer,Thank you for your continued support of Xiaomi MiMo. The Xiaomi MiMo team has been continuously optimizing our inference performance, and we are pleased to pass these cost savings directly to you.Significant Price Reductions for MiMo-V2.5 (Up to 99% Off)New pricing will be effective from May 27, 2026, 00:00 (GMT+8). We have simplified our structure by removing length-based tiers. V2 model pricing remains unchanged and these models will be deprecated soon. We advise migrating to the new models as soon as possible.MiMo-V2.5-Pro (Per 1M tokens)Input (Cache Hit): $0.0036Input (Cache Miss): $0.435Output: $0.87MiMo-V2.5 (Per 1M tokens)Input (Cache Hit): $0.0028Input (Cache Miss): $0.14Output: $0.28MiMo-V2.5-TTS Series: Remains Free for a Limited TimeView API Pricing →Token Plan: More Value, Same PriceIncreased Credits: Usage quotas for the MiMo-V2.5 series have increased by 3x-5x. We have also optimized our measurement metrics across cached input, input, and output for greater transparency.Full Credit Reset: For all active plans (Including the Token Plan obtained through participation in the “MiMo Orbit” 100T Token Incentive Program, covering exclusive benefits of the Apache Software Foundation), your current credit usage will be completely reset, while your existing validity period remains unchanged.One More Thing: For historical paying users whose token plans have expired, we have also prepared surprise gifts, which will be announced in the coming week. Please stay tuned.Learn More About Token Plan →”MiMo Orbit” 100T Token Incentive Program ConcludesSince launching on April 28, 2026, the program received overwhelming global participation. As of May 26, 2026, 16:08 (GMT+8), the full 100T tokens have been claimed. The program has officially concluded. Thank you to all participating developers!Learn More →Technical Insights: Inference OptimizationThese price adjustments are the result of our team’s relentless optimization of the inference system:By leveraging SGLang HiCache and providing full support for Sliding Window Attention (SWA), we have reduced data movement across multi-level storage (GPU VRAM, CPU RAM, and SSD) to nearly 1/7 of previous levels, while increasing the cacheable token capacity by 5x. These advancements significantly improve cache hit rates and overall efficiency. Furthermore, by refining our expert-parallelism schemes and input-length bucketing strategies, we have boosted cluster throughput, allowing us to lower per-token costs without compromising service quality.Best regards, Xiaomi MiMo API Platform Team |
| Xiaomi MiMo API 开放平台团队2026 年 5 月 |
意见反馈