Xiaomi MiMo-V2.5 Series Large Models Launch Public Beta

Today, the Xiaomi MiMo-V2.5 series models officially launch their public beta.

The Xiaomi MiMo-V2.5 series includes MiMo-V2.5, V2.5-Pro, V2.5-TTS Series, and V2.5-ASR.

Featuring stronger reasoning capabilities, more stable Agent performance, extended context windows, improved instruction following and fuzzy instruction understanding, plus enhanced full-modality perception and comprehension—this represents a comprehensive leap from "usable" to "truly excellent."

At the same time, we have optimized our Token Plan pricing scheme—making the world's top-tier models more accessible than ever.

MiMo-V2.5-Pro: Stronger Agents, Extended Focus

MiMo-V2.5-Pro is our most powerful model to date. In dimensions such as general agent capabilities, complex software engineering, and long-horizon tasks, it can now compete head-to-head with global top-tier Agent models (Claude Opus 4.6, GPT-5.4), achieving an all-around upgrade compared to the previous generation MiMo-V2-Pro.

In internal testing, MiMo-V2.5-Pro demonstrated an intelligence level that makes us rethink human-model collaboration: paired with an appropriate execution framework, it can stably complete long-horizon tasks involving nearly a thousand tool calls in a single run. Its instruction-following capability in agent scenarios has also significantly improved—precisely capturing implicit requirements within context while maintaining logical consistency over ultra-long periods. Thus, MiMo-V2.5-Pro can now undertake truly serious professional work with higher confidence.

Image

Built for More Complex Tasks

MiMo-V2.5-Pro is designed for higher-difficulty, more complex task objectives. We assigned it tasks that would typically take human experts days or even weeks to complete, allowing it to independently execute long-horizon workflows while maintaining extremely high quality. Here are its delivered results:

  • Implementing a Complete SysY Compiler in Rust

This task originates from Peking University's "Compiler Principles" course project, requiring the model to implement a complete SysY compiler from scratch using Rust: lexical analyzer, syntax analyzer, AST, Koopa IR code generation, RISC-V assembly backend, and performance optimization. For reference, it typically takes Peking University undergraduates several weeks to complete this project, yet MiMo-V2.5-Pro finished all work in just 4.3 hours through 672 tool calls, achieving a perfect score of 233/233 on hidden test sets, demonstrating extremely high productivity value.

Image

It did not fall into brute-force trial and error but systematically built the entire compiler layer by layer: first establishing the complete pipeline skeleton, then tackling each layer sequentially—achieving full marks in Koopa IR (110/110), RISC-V backend (103/103), and performance optimization (20/20). The first compilation passed 137/233 tests, with a 59% cold-start pass rate, meaning the architecture was correct even before running any tests. At iteration 512, a refactor caused two test points in lv9/riscv to regress; the model self-diagnosed, recovered, and continued progressing.

Long-horizon tasks reward precisely this kind of structured, self-correcting work discipline.

  • Developing a Video Editor

With just a few simple instructions—"Build a video editor web application"—MiMo-V2.5-Pro delivered a functional web app featuring multi-track timelines, clip trimming, cross-fading, audio mixing, and export workflows. The final build comprised 8,192 lines of code, completed through 1,868 tool calls over 11.5 hours of autonomous work.

若影片無法播放,請改看 來源頁

MiMo-V2.5: Next-Level Full-Modality Agent with Million-Token Context

MiMo-V2.5 is a native full-modality large model born for Agent scenarios, capable of seeing, hearing, reading, and translating understanding into action simultaneously.

This time, MiMo-V2.5 brings two key upgrades:

Agent capabilities comprehensively surpass MiMo-V2-Pro

In authoritative Agent evaluations such as Claw-Eval, MiMo-V2.5 exceeds MiMo-V2-Pro performance levels, handling daily simple tasks effectively while reducing API costs by approximately 50%.

Multi-modality perception comprehensively surpasses MiMo-V2-Omni

Capabilities in cross-modal reasoning, video understanding, and chart analysis have improved, approaching or even surpassing top-tier proprietary industry models in evaluations like VideoMME, CharXiv, and MMMU-Pro.

Image

MiMo-V2.5 Full Series: Higher Token Efficiency

The entire MiMo-V2.5 series has been optimized for token efficiency, accomplishing more with fewer tokens.

When achieving identical scores on the Agent benchmark ClawEval:

  • MiMo-V2.5-Pro saves 42% more tokens compared to Kimi K2.6

  • MiMo-V2.5 saves 50% more tokens compared to Muse Spark

Image

MiMo-V2.5 Full Series: How to Use Them Together?

  • MiMo-V2.5-Pro is tailored for long and difficult Agent tasks, while MiMo-V2.5 covers the vast majority of general Agent scenarios

  • MiMo-V2.5 supports native full-modality Agent capabilities, covering images, audio, and video

  • MiMo-V2.5 offers higher average inference speeds, enabling faster responses for latency-sensitive tasks

Image

Token Plan Refresh and Upgrade

We have made several substantial optimizations to the Token Plan tailored for your needs:

Credits Rate Update: More Affordable

  • MiMo-V2.5: 1x (1 Token consumed = 1 Credit)

  • MiMo-V2.5-Pro: 2x (1 Token consumed = 2 Credits)

The previous billing method of 1 Token = 4 Credits has been cancelled. From now on, the Token Plan no longer differentiates Credit multipliers between 256k and 1M context windows.

Nighttime Exclusive Discount Rate

From 00:00 to 08:00 Beijing Time daily, all model Credits consumption rates receive an additional 20% discount on top of existing rates.

Auto-Renewal Discounts

A new "Continuous Monthly Subscription" mode has been added: existing users enabling auto-renewal enjoy 30% off the following month, while new users enjoy 23% off the following month, limited to one-time use.

A new "Annual Subscription" cycle has been added: one-time subscription enjoys 12% off the entire year, without stacking with first-purchase or auto-renewal discounts.

Launch Bonus: Full Reset of Token Plan User Credits

All existing Token Plan users (as of 22:00 Beijing Time, April 22) will have their Credits balance fully reset to zero, restarting calculations anew.

Xiaomi MiMo helps you start from scratch and unleash your creativity to the fullest!

Note: This launch bonus only resets the Credits quota; subscription timing remains unchanged. Validity periods for already purchased plans remain unaffected.

Image

Coming Soon to Open Source

MiMo-V2.5-Pro and MiMo-V2.5 models will soon be open-sourced globally. Stay tuned.

Official API Access
Xiaomi MiMo Open Platform: https://platform.xiaomimimo.com
Token Plan Subscription: https://platform.xiaomimimo.com/#/token-plan

Product Experience
Xiaomi MiMo Studio: https://aistudio.xiaomimimo.com/#/c

Image

https://platform.xiaomimimo.com

https://platform.xiaomimimo.com/#/token-plan

https://aistudio.xiaomimimo.com/#/c

Related Articles

分享網址
AINews·AI 新聞聚合平台
© 2026 AINews. All rights reserved.