Xiaomi Unveils 1.02‑Trillion‑Parameter MiMo 2.5 Model – Token Grant Guide and Real‑World Benchmarks

MiMo 2.5 model family

Three models were released simultaneously: MiMo‑V2.5‑Pro (1.02 T total parameters, 42 B active), MiMo‑V2.5 (310 B total, 15 B active), and MiMo‑V2.5‑ASR (8 B, speech‑only). All use a Mixture‑of‑Experts (MoE) architecture, activating only a fraction of parameters per inference, which enables a 1 M‑token context for text, image, video, and audio.

Token‑grant program

Developers can apply for a free subscription at https://100t.xiaomimimo.com/. Two tiers are offered: “Max” with 1.6 B tokens and “Standard” with 200 M tokens. After applying, an email is sent the same day and the account must be registered on the open platform using the same email address.

Model specifications

MiMo‑V2.5‑Pro : 1.02 T total, 42 B active, 1 M‑token context, MIT license.

MiMo‑V2.5 : 310 B total, 15 B active, 1 M‑token context, MIT license.

MiMo‑V2.5‑ASR : 8 B total, speech‑only, Apache 2.0 license.

Benchmark performance (selected results)

GPQA Diamond: 66.7 (MiMo‑Pro) vs 85.7 (GPT‑5 High) – lower than top‑tier models.

GSM8K: 99.6 (MiMo‑Pro) – highest among listed models.

AIME 2025: 94.1 (MiMo‑Flash) vs 95.0 (Gemini 3.0 Pro).

MMLU‑Pro: 68.5 (MiMo‑Pro) vs 90.1 (Gemini 3.0 Pro).

Code‑generation and agent benchmarks

SWE‑Bench Verified: 78.9 (MiMo‑Pro) > 74.9 (GPT‑5 High) and > 76.2 (Gemini 3.0 Pro).

SWE‑Bench Pro: 57.2 (MiMo‑Pro) – only reported score for this suite.

TerminalBench 2: 68.4 (MiMo‑Pro) vs 54.2 (Gemini 3.0 Pro) and 42.8 (Claude Sonnet 4.5).

SWE‑Bench Multilingual: 71.7 (MiMo‑Flash) vs 55.3 (Gemini 3.0 Pro).

Real‑world usage examples

Using the same prompts that built a personal blog with NextJS (frontend) and Spring Boot (backend) via DeepSeek‑V4, MiMo‑V2.5‑Pro produced comparable front‑end results.

For UI‑to‑website generation, DeepSeek failed because it lacks multimodal support. MiMo‑V2.5‑Pro processed the image and generated a partial layout, but reconstruction quality was below 50 % and required multiple iterations.

Comparisons with Claude Code + MiMo‑V2.5‑Pro and GPT‑5.5 High showed a noticeable gap in visual‑to‑code fidelity, indicating MiMo still trails the latest multimodal models.

Token consumption and cost

During testing, roughly 50 M tokens were consumed within an hour, approaching the 200 M‑token limit of the Standard plan. By contrast, DeepSeek‑V4 used less than 40 M tokens for similar workloads, suggesting higher token cost for MiMo at current pricing.

Conclusion

MiMo 2.5 places Xiaomi among the few providers of open‑source trillion‑parameter models (alongside DeepSeek, Kimi, Qwen). The MoE design, 1 M‑token context, and multimodal capabilities make the models competitive for developers, despite higher token consumption and still‑evolving visual‑code generation performance.

Code example

最后欢迎
加入苏三的星球
，你将获得：智能商城推荐AI Agent、智能代码审查AI Agent、
智能天气播报AI Agent、
SaaS点餐系统（DDD+多租户）、100万QPS短链系统（超过并发）、复杂的
商城微服务系统（分布式）、
苏三
商城系统、苏三AI项目、刷题吧小程序、
秒杀系统
、
代码生成工具等12个实战项目的
源代码
、
开发教程
和
技术答疑
。
系统设计、性能优化、技术选型、底层原理、Spring源码解读、工作经验分享、痛点问题
、面试八股文
等多个优质专栏。
推荐12个牛逼的SpringBoot项目
推荐一个牛逼的RAG + MCP系统
还有1V1免费修改简历、技术答疑、职业规划、送书活动、技术交流。
扫描下方二维码，可以加入星球：
原价159，券后仅需129。
数量有限，先到先得。
目前星球已经更新了
6100+
篇优质内容，还在持续爆肝中.....
星球已经被
官方推荐了3次
，收到了小伙伴们的一致好评。戳我加入学习，已有
2600+
小伙伴加入学习