In a surprise late-night release, Xiaomi has officially launched its highly anticipated self-developed MiMo-V2 series of large Xiaomi AI models. Comprising three specialized tiers—MiMo-V2-Pro, MiMo-V2-Omni, and MiMo-V2-TTS—this new lineup marks Xiaomi’s aggressive push into the “Agent Era” of artificial intelligence.
While native ecosystem integrations within apps like Xiaomi Browser and Kingsoft Office are currently exclusive to the Chinese market, the models are entirely browser-based, allowing developers and enthusiasts worldwide to explore their capabilities right now via the official API website or directly Xiaomi MiMo Studio.
Core Content: The Technical Breakdown of the MiMo-V2 Series
Xiaomi’s latest release isn’t just a minor step forward; the benchmark data proves these models are built to compete with the heavyweights of the AI industry.
Xiaomi MiMo-V2-Pro: The Heavy-Duty Agent
Designed for high-intensity, complex workflows without human intervention, the flagship Pro model is a powerhouse for logic reasoning and task planning.
-
Technical Specs: It boasts a massive 1 Trillion (1T) total parameters with 42 Billion (42B) activated during inference. Utilizing an innovative mixed-attention architecture, it supports an ultra-long context window of 1M tokens.
-
Benchmark Dominance: Tested under the mysterious codename “Hunter Alpha,” the model recently broke the 1T token usage mark during testing. On the rigorous Claw-Eval benchmark, MiMo-V2-Pro (Hunter Alpha) achieved an impressive average score of 75.7, placing it comfortably in the top three globally, directly trailing Anthropic’s Claude Opus 4.6. Furthermore, on the Artificial Analysis Intelligence Index, it secured an index score of 49, ranking it second in China and eighth globally—surpassing competitors like Grok 4.20 and Gemini 3 Flash.
-
Code & Execution: Internal engineer reviews indicate its coding capabilities—system design, workflow orchestration, and elegant code generation—feel incredibly close to Claude Opus 4.6, but at a fraction of the API cost.
Xiaomi MiMo-V2-Omni: The Full-Modal Multimodal Base
The Omni model is Xiaomi’s answer to seamless cross-modality understanding, natively processing image, video, audio, and text inputs.
-
Benchmark Triumphs: Tested under the codename “Healer Alpha,” it dominated the PinchBench leaderboard. In direct comparisons, MiMo-V2-Omni outperformed heavy hitters like Gemini 3 Pro and Claude Opus 4.6 in key areas:
-
Speech Reasoning (BigBench Audio): Scored an astounding 94.0.
-
Audio Understanding (MMAU-Pro): Topped the charts with 69.4.
-
Video Future Event Forecast (FutureOmni): Led the pack with 66.7.
-
-
Real-World Application: It can autonomously develop and execute plans across different modalities, remediating policies in real-time if it encounters anomalies.
Xiaomi MiMo-V2-TTS: Giving the Agent a Soul
To complement the reasoning models, Xiaomi introduced a state-of-the-art speech synthesis model based on a self-developed Audio Tokenizer and multi-codebook joint modeling.
-
Hyper-Realistic Control: Trained on hundreds of millions of hours of audio data and refined via multi-dimensional reinforcement learning, the TTS model allows for precise, multi-granular emotional control. It can transition emotions and tone mid-sentence, sing with accurate pitch, and natively synthesize various regional dialects (including Sichuan, Henan, Cantonese, and Taiwanese accents).
Platform Availability and API Pricing
As a browser-based architecture, the barrier to entry is extremely low. While the native integrations (like the Kingsoft WebOffice ecosystem bridging MiMo with Word, Excel, PPT, and PDF) are targeted at the Chinese market initially, the models are globally accessible.
Xiaomi has made the API available immediately at platform.xiaomimimo.com, priced highly competitively in USD:
-
MiMo-V2-Pro:
-
Up to 256K Context: $1.00 / 1M Input tokens | $3.00 / 1M Output tokens
-
Up to 1M Context: $2.00 / 1M Input tokens | $6.00 / 1M Output tokens
-
-
MiMo-V2-Omni:
-
Up to 256K Context: $0.40 / 1M Input tokens | $2.00 / 1M Output tokens
-
Note: For a limited time, developers can test these models for free for one week via agent frameworks like OpenClaw, OpenCode, KiloCode, Blackbox, and Cline.
Xiaomi’s Long-Term Strategy
This late-night release solidifies Xiaomi’s commitment to building a formidable software ecosystem. By offering Claude Opus 4.6-level performance at roughly 20% of the cost, Xiaomi is drastically lowering the threshold for cutting-edge AI implementation.
