In the early hours of March 19, 2026, Xiaomi's large‑model team announced on its official website and multiple platforms the official release of MiMo‑V2‑Pro, MiMo‑V2‑Omni and MiMo‑V2‑TTS—three self‑developed large models. This is not only another technical iteration of the MiMo series but also a milestone in Xiaomi's explicit decision to "go all‑in on the Agent era". The three models are positioned as a flagship base, a multimodal base and a speech‑synthesis model respectively, together forming a complete Agent capability chain from perception and reasoning to execution and expression. They have been seamlessly integrated into ecosystems such as Xiaomi miclaw, MiMo Studio, Kingsoft WPS Lingxi and Xiaomi Browser, and are available for a limited one‑week free trial via frameworks like OpenClaw and OpenCode. Lei Jun wrote on Weibo that Xiaomi's AI R&D and capital investment will exceed RMB 16 billion this year and that actual progress "may be much faster than people see".
MiMo‑V2‑Pro: Trillion‑parameter Agent flagship base, eighth globally and second in China
MiMo‑V2‑Pro is the core flagship released this time, designed for high‑intensity real‑world Agent workloads. Total parameters exceed 1 trillion (42B activated parameters); it adopts an innovative hybrid‑attention architecture and natively supports 1M long context. During a previous anonymous rollout on OpenRouter under the codename "Hunter Alpha", calls surpassed 1 trillion tokens, and it ranks eighth globally and second in China on the Artificial Analysis global large‑model comprehensive intelligence ranking (behind closed‑source leaders such as Gemini 3.1 Pro Preview and ahead of xAI Grok).
The model focuses on breakthroughs in generalization from "Coding to Claw": it performs strongly on complex, multi‑step, long‑horizon Agent tasks and already supports tool invocation, function execution and GUI operations across the full stack. API pricing uses tiered billing: within 256K context input $1 /M tokens and output $3 /M tokens; within 1M context input $2 and output $6, offering significant cost‑effectiveness. It is deeply integrated into mainstream Agent frameworks such as OpenClaw, providing developers with a reliable "intelligent brain".
MiMo‑V2‑Omni: Multimodal Agent base that sees, hears and acts
MiMo‑V2‑Omni is Xiaomi's first base model built from the ground up to fuse text, vision and audio for multimodal interaction and execution scenarios. During prior anonymous testing under codename "Healer Alpha", it achieved the top average score on the OpenClaw benchmark PinchBench.
The core advantage lies in the "deep binding of perception and action":
- Audio understanding: supports environmental sound classification, multi‑speaker separation, audio‑video joint inference and understanding of audio longer than 10 hours, with overall performance surpassing Gemini 3 Pro.
- Image/video understanding: outperforms Claude Opus 4.6 on multidisciplinary visual reasoning and complex chart analysis, approaches Gemini 3 Pro levels, and natively provides contextual awareness and future inference capabilities.
- Execution capability: seamlessly supports tool calls, function execution and GUI operations, and can directly manipulate interfaces, greatly lowering the barrier to deploying multimodal Agents.
API is open: for 256K context pricing is CNY 0.4 /M tokens input and CNY 2 /M tokens output. The release of Omni marks Xiaomi's shift from "language understanding" to "multimodal execution" for Agents.
MiMo‑V2‑TTS: New benchmark in speech synthesis, speaks, sings and handles dialects with expressive control
MiMo‑V2‑TTS is Xiaomi's in‑house speech‑synthesis large model, built on a proprietary Audio Tokenizer and a multi‑codebook speech‑text joint modeling architecture, pretrained on over 100 million hours of speech data and further refined with reinforcement learning. Highlights include:
- Highly controllable multi‑granularity style control: can adjust from overall intonation to local emotion precisely, even achieving tone shifts within a single sentence.
- Dialect and expressiveness: supports natural pronunciations of Northeast Mandarin, Sichuanese, Henan dialect, Cantonese and Taiwan accent, can perform role‑play style renditions and high‑quality singing synthesis with accurate pitch and rhythm.
- Future plans: to deeply integrate with MiMo‑V2‑Omni so Agents not only "understand the world" but can "tell the world" using expressive human‑like voices.
Ecosystem deployment and developer benefits: a full cloud‑to‑device closed loop
The three models are simultaneously available on Xiaomi miclaw, MiMo Studio, Kingsoft WPS Lingxi, Xiaomi Browser and other platforms, and have achieved framework‑level integration with five major Agent frameworks: OpenClaw, OpenCode, KiloCode, Blackbox and Cline. Developers can try them free for a limited week; WPS Lingxi has already integrated MiMo‑V2‑Pro, allowing users to assign tasks to "Lingxi Claw" for more efficient office workflows.
Xiaomi emphasizes that the MiMo series is accelerating deep integration with devices including phones (Xiaomi 17 series), cars (SU7) and whole‑home smart systems (Miloco), enabling cloud‑edge‑device collaboration and privacy protection. This aligns closely with Xiaomi's "people‑car‑home full‑ecosystem" strategy and provides foundational support for mass production of humanoid robots in 2026.
Industry significance: Xiaomi moves from "catching up" to "defining Agent rules"
In the intensely competitive global AI landscape of 2026, Xiaomi's triple release demonstrates a clear Agent strategy: Pro handles high‑intensity cognition and planning, Omni handles multimodal perception and execution, and TTS handles natural expression. Together they build a complete Agent loop that significantly lowers deployment barriers.
Compared with pure cloud models, Xiaomi's advantage lies in massive real‑user data and its IoT ecosystem, allowing models to truly serve the "real world". Reports from institutions such as Goldman Sachs believe Xiaomi has entered a period of AI results realization, and its vertically integrated advantages will drive long‑term value reshaping.
Conclusion: Xiaomi is in position for the Agent era
From the trillion‑parameter flagship MiMo‑V2‑Pro to Omni's multimodal execution and TTS's expressive capabilities, Xiaomi delivers a solid answer for the Agent era with three models. This is more than a tech release; it is a declaration of Xiaomi's transformation from a hardware giant to an AI full‑stack ecosystem player. Driven by computing power and data, Xiaomi is pragmatically defining China's unique competitiveness in the Agent era. As more ecosystem partners connect and open‑source efforts advance, the MiMo series may become the preferred base for global developers building agents.
▶️ Enjoy this article? Creating this is hard work!
Follow our YouTube channel:
AI Rosetta
Your likes, subscriptions and shares are our greatest support! 🙏
More AI news at: https://www.airosetta.com