Xiaomi's MiMo-V2.5: A Game Changer in AI Technology
Xiaomi on Wednesday released the MiMo-V2.5 series of AI models, a rapid follow-up to its MiMo-V2-Pro that debuted just five weeks ago. The new lineup includes MiMo-V2.5-Pro, the company's flagship foundation model, alongside MiMo-V2.5, a natively multimodal model that processes images, video, audio, and text within a single architecture.
The V2.5 series entered public beta on April 22, with Xiaomi signaling plans to open-source the models. The company also updated its TokenPlan subscription system to cover all V2.5 models, maintaining the four-tier pricing structure it introduced in early April.
Building on V2-Pro's Foundation
MiMo-V2-Pro landed on March 18 as a text-only model with over one trillion total parameters and 42 billion active via a mixture-of-experts architecture. It quickly earned attention after appearing anonymously on OpenRouter under a codename, leading many to speculate it was DeepSeek V4 before Xiaomi claimed credit.
The V2.5-Pro retains the same trillion-parameter, 42-billion-active architecture with a one-million-token context window. On the agentic side, Xiaomi says it matches Claude Opus 4.6 in demanding scenarios and ranks among the top models on benchmarks including ClawEval, GDPVal, and SWE-bench Pro.
Multimodal Model at Lower Cost
The standout addition is MiMo-V2.5, the omnimodal variant. Unlike V2-Pro's text-only design, MiMo-V2.5 natively understands images, video, audio, and text, enabling cross-modal perception and long-range reasoning. According to OpenRouter, MiMo-V2.5 delivers "Pro-level agentic performance at roughly half the inference cost" — priced at $0.40 per million input tokens and $2 per million output tokens, compared to V2-Pro's $1 and $3 respectively.
Xiaomi positions the multimodal model as suited for real-time perception-action applications, including robotics and smart home integration — consistent with its broader "Human x Car x Home" ecosystem strategy. The model also ships with built-in agentic capabilities for browsing, reasoning, and task execution.
Ecosystem and Pricing Push
Alongside the model release, Xiaomi expanded its TokenPlan to support monthly and annual subscriptions across the full V2.5 and V2 series. The four tiers range from a Lite plan for individual developers to a Max plan for heavy users, with credits drawn down based on token consumption. Xiaomi reported that its MiMo platform processed over one trillion tokens on its first day of commercial availability in early April.
The V2.5 series is compatible with mainstream development toolchains including OpenCode, OpenClaw, and Claude Code, reinforcing Xiaomi's bid to establish MiMo as infrastructure for the growing AI agent ecosystem.
