Xiaomi MiMo-V2.5 Series Large Model Launches Public Beta

Today, the Xiaomi MiMo-V2.5 series of models officially launched its public beta.

The Xiaomi MiMo-V2.5 series includes mimo-v2.5, v2.5-pro, V2.5-TTS Series, and v2.5-asr.

Stronger reasoning, more stable agents, longer context, stronger instruction following and understanding of ambiguous instructions, better full-modal perception and understanding —— this is a comprehensive leap from "usable" to "user-friendly".

Meanwhile, we have also optimized the Token Plan pricing plan —— making the world's top-notch models easily accessible.

mimo-v2.5-pro: Stronger Agent, Longer Focus

mimo-v2.5-pro is our most powerful model to date. In dimensions such as general agent capabilities, complex software engineering, and long-range tasks, it can already compete head-on with the world's top Agent models (Claude Opus 4.6, GPT-5.4), achieving a comprehensive leap compared to the previous generation mimo-v2-pro.

During internal testing, the intelligence level demonstrated by mimo-v2.5-pro has made us rethink the way humans and models collaborate: when paired with a suitable operating framework, it can stably complete long-range tasks involving nearly a thousand rounds of tool calls in a single instance, and its instruction-following ability in the agent scenario has also significantly improved - it can accurately capture implicit requirements in the context and maintain logical consistency over an extremely long period. By now, mimo-v2.5-pro can already undertake truly serious professional work with a higher confidence level.

Designed for more complex tasks

mimo-v2.5-pro is designed for more challenging and complex task objectives. We assign tasks that would take human experts days or even weeks to complete to it, allowing it to independently complete long-term processes while still maintaining extremely high quality. The following are the results it has delivered:

Implement a complete SysY compiler in Rust

This task originated from the Compilation Principles course project at Peking University, requiring the model to implement a complete SysY compiler from scratch in Rust, including a lexical analyzer, a syntax analyzer, AST, Koopa IR code generation, RISC-V assembly backend, and performance optimization. For reference, undergraduate students at Peking University usually take several weeks to complete this project, while mimo-v2.5-pro only took 4.3 hours, completed all tasks after 672 tool calls, and achieved a perfect score of 233/233 on the hidden test set, demonstrating extremely high productivity value.

Instead of getting stuck in brute-force trial-and-error, it builds the entire compiler layer by layer: first constructing the complete pipeline framework, then tackling each layer one by one - Koopa IR achieved a perfect score (110/110), the RISC-V backend achieved a perfect score (103/103), and Performance optimization achieved a perfect score (20/20). The first compilation passed 137/233 , with a cold start pass rate of 59%, which means that the architecture was already correct before running any tests. In the 512th round, a refactoring caused lv9/riscv to regress by two test points; the model self-diagnosed, recovered, and continued to progress.

The long-range task rewards precisely this structured and self-correcting work discipline.

Develop a video editor

With just a few simple instructions - "Build a video editor web application" - mimo-v2.5-pro delivered a runnable web application: featuring multi-track timeline, clip trimming, cross-fading, audio mixing, and export processes. The final built codebase amounts to 8,192 lines, involves 1,868 tool invocations, and was completed in 11.5 hours of autonomous work.

mimo-v2.5 is a native full-modal large model designed for Agent scenarios, capable of seeing, hearing, and reading simultaneously, and translating understanding into action.

This time, mimo-v2.5 brings a key upgrade:

Agent capabilities comprehensively surpass mimo-v2-pro

In authoritative Agent evaluations such as Claw-Eval, mimo-v2.5 surpasses the level of mimo-v2-pro, is capable of handling daily simple tasks, and at the same time reduces API costs by approximately 50%.

MultiModal Machine Learning perception comprehensively surpasses mimo-v2-omni

Capabilities such as cross-modal reasoning, video understanding, and chart analysis have been enhanced, approaching and even surpassing industry-leading closed-source models in evaluations such as VideoMME, CharXiv, and MMMU-Pro.

MiMo-V2.5 Full Series: Higher Token Efficiency

The entire MiMo-V2.5 series is optimized for Token efficiency, doing more with fewer Tokens.

When achieving the same score on the Agent benchmark list ClawEval:

mimo-v2.5-pro saves 42% Token compared to Kimi K2.6
mimo-v2.5 saves 50% of tokens compared to Muse Spark

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

mimo-v2.5-pro is specifically designed for long and complex Agent tasks, while mimo-v2.5 covers most general Agent scenarios
mimo-v2.5 supports native full-modal Agent capabilities, covering images, audio, and video
mimo-v2.5 has a higher average inference speed and can respond more quickly to latency-sensitive tasks

Token Plan Upgraded and Refreshed

We have made several substantial optimizations suitable for you regarding the Token Plan:

Credits rate updated, more favorable

mimo-v2.5：1x（use 1 Token = 1 Credit）
mimo-v2.5-pro： 2x（use 1 Token = 2 Credits）

Cancel the billing method of 1 Token = 4 Credits. From now on, the Token Plan will no longer distinguish the Credit multiplier for 256k and 1M context windows.

Exclusive Nighttime Discount Rate

From 00:00 to 08:00 Beijing Time every day, the consumption rate of Credits for all modelswill be further discounted by 20% on top of the original rate.

Enjoy discounts with auto-renewal

A new "Continuous Monthly Subscription" model has been added. Existing users who activate auto-renewal will enjoy a 30% discount on the next month's subscription, while new users will enjoy a 23% discount on the next month's subscription, both limited to one time.

A new "Annual" subscription cycle has been added. Subscribing once will enjoy an 12% discount for the whole year, and no longer be combined with the first purchase/auto-renewal discount.

Online Benefit: Token Plan users' Credits will be fully reset

All users who have purchased the Token Plan (as of 22:00 on April 22, Beijing Time) will have their Credits quota fully reset to zero, and the calculation will start anew.

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

Note: This online welfare only resets the Credits limit, does not reset the package timing, and the validity period of purchased packages remains unchanged.

is about to be open-sourced

mimo-v2.5-pro and mimo-v2.5 models are about to be globally open-sourced. Stay tuned!

Update Time May 28, 2026