Xiaomis Stealth AI Stunt Precedes Major $83 Billion Investment Pledge

Xiaomi Unveils Ambitious AI Investment and Model Suite, Following Stealth API Launch That Fooled Industry

A week of intense speculation within the artificial intelligence developer community was put to rest as Chinese technology giant Xiaomi officially claimed responsibility for two anonymously released, high-performing AI models. The revelation coincided with a major corporate announcement of a sweeping, multi-billion dollar investment into AI over the next three years, signaling the smartphone and consumer electronics maker's determined push to become a leading force in the foundational model arena.

The intrigue began when two unnamed models, codenamed "Hunter Alpha" and "Healer Alpha," appeared without fanfare on the popular API aggregation platform OpenRouter. Despite the complete absence of official marketing, their usage rates began climbing at an unusual pace. Hunter Alpha repeatedly topped the platform's daily leaderboard, with cumulative usage surpassing 1 trillion tokens, sparking widespread discussion. The leading theory within developer forums pointed towards DeepSeek, a prominent Chinese AI research organization, with many speculating these were internal test versions of a presumed "DeepSeek V4."

The mystery deepened when Peter Steinberger, founder of the OpenClaw agent framework, publicly inquired about the models' origin on social media platform X, further fueling community curiosity. The puzzle was solved when Xiaomi officially confirmed that both Hunter Alpha and Healer Alpha were early internal test versions of its MiMo large language model series. Luo Fuli, head of Xiaomi's MiMo large model team, publicly claimed the models on X. In a notable twist, Luo is a former researcher at DeepSeek, meaning a DeepSeek alumnus at Xiaomi built models compelling enough to be mistaken for DeepSeek's own.

Strategic Bet: A 600-Billion-Yuan Commitment The model unveiling was framed within a much broader strategic context. At a recent product launch event, Lei Jun, founder, chairman, and CEO of Xiaomi Group, declared that the company plans to invest over 600 billion yuan (approximately $83 billion USD) in the AI field over the next three years. This staggering financial commitment underscores AI as the central pillar of Xiaomi's future. Lei also announced the company's first "Lobster" series smartphone, the Xiaomi Miclaw, which has entered closed beta testing. The device is designed to deeply integrate the MiMo large model across Xiaomi's operating system and its "Human x Car x Home" full ecosystem.

The newly announced MiMo-V2 model family, which includes the anonymously tested versions, is explicitly engineered to advance AI from conversational ability to task completion. Xiaomi introduced three interlinked models, each with a distinct focus within an agentic framework.

MiMo-V2-Pro: The High-Performance, Value-Priced Workhorse Positioned as the flagship text-based foundation model for high-intensity agent workloads, MiMo-V2-Pro specializes in reasoning, planning, and tool calling. Its total parameter count exceeds 1 trillion, with 42 billion activated parameters—roughly three times larger than its predecessor, MiMo-V2-Flash. The model maintains inference efficiency through an innovative Hybrid Attention architecture and a lightweight Multi Token Prediction layer. It also supports an ultra-long context window of 1 million tokens, a structural advantage for complex, multi-step agent tasks.

On the comprehensive global model benchmark Artificial Analysis, MiMo-V2-Pro currently ranks eighth worldwide and second in China. Beyond benchmarks, Xiaomi emphasizes "practical performance." The company states that in dimensions like Coding Agent, General Agent, and Tool Use, MiMo-V2-Pro is on par with models like Claude Sonnet 4.6. Internal engineer evaluations suggest its code engineering capability is接近 (approaching) Claude Opus 4.6, with superior system design ability and more elegant code style. Data from the Hunter Alpha stealth test provided market validation, as the highest usage categories were predominantly programming-specific tools.

Perhaps its most disruptive feature is its pricing strategy. Xiaomi has priced MiMo-V2-Pro's API at approximately one-fifth the cost of comparable competitors. For contexts within 256K tokens, pricing is set at $1 per million tokens for input and $3 for output. For the full 1M context, it is $2 for input and $6 for output. To accelerate developer adoption, Xiaomi is partnering with five major agent framework teams—OpenClaw, OpenCode, KiloCode, Blackbox, and Cline—to offer a one-week limited free API access period. The model is now openly available via the MiMo platform.

MiMo-V2-Omni: A Multimodal Agent with Eyes, Ears, and Hands While MiMo-V2-Pro serves as the "brain," MiMo-V2-Omni's ambition is to equip that brain with perceptual and acting capabilities. It is Xiaomi's first foundation model to natively unify perception and action at the architectural level, deeply integrating text, vision, and audio from the ground up.

Its audio understanding capability is highlighted as a key differentiator, supporting over 10 hours of continuous long-audio comprehension and outperforming Google's Gemini 3 Pro in complex scenarios. In image understanding, it is reported to surpass Claude Opus 4.6 in multi-disciplinary visual reasoning and complex chart analysis. For video, it supports native audio-visual joint input, providing a genuine multimodal advantage.

In practical agent demonstrations, MiMo-V2-Omni showcased impressive end-to-end task completion. Integrated with the OpenClaw framework, it can autonomously control a browser to perform tasks like researching product reviews on social media, comparing prices across e-commerce sites, negotiating with customer service for discounts, and finalizing a purchase. Another demonstration showed the model taking a verbal command to "make an intro video with techy sound effects and post it to TikTok," subsequently executing the entire workflow, including troubleshooting a font rendering error and confirming the video's successful upload.

The model has also been integrated into office productivity software through a partnership with Kingsoft Office, enabling direct generation of Word documents, structured Excel sheets, and formatted PPT presentations within WPS. MiMo-V2-Omni's API is now open, priced at $0.4 per million tokens for input and $2 for output within a 256K context window.

MiMo-V2-TTS: Giving Agents a Voice Completing the agent stack is MiMo-V2-TTS, a speech synthesis model designed to give AI agents a natural, expressive voice. Built on Xiaomi's proprietary Audio Tokenizer and a multi-codebook joint speech-text modeling architecture, it was pre-trained on "hundreds of millions of hours" of speech data. This scale allows it to cover a vast range of speaking styles, accents, and scenarios. The model supports multi-granularity control, from setting an overall tone to managing intra-sentence emotional shifts and emphasis. It can automatically interpret punctuation and formatting cues to produce natural-sounding speech without manual intervention.

Analysis: Aggressive Posturing in a Crowded Field Xiaomi's coordinated reveal—first seeding high-quality models anonymously to generate organic buzz, then following with a major investment announcement and detailed product launch—demonstrates a sophisticated go-to-market strategy for its AI division. The recruitment of top talent like Luo Fuli from leading AI labs is a critical component of its technical execution.

The company's approach is distinctly pragmatic and market-oriented. By pricing its flagship text model aggressively, Xiaomi clearly aims to quickly capture developer mindshare and build an ecosystem, a proven tactic in its consumer hardware business. The focus on "agentic" models that "complete tasks," rather than just converse, aligns with the industry's shift towards actionable AI. Furthermore, the deep integration of these models into its own ecosystem—from smartphones to office software—provides a built-in deployment channel and differentiates its strategy from pure-play AI model companies.

However, the AI foundation model space is intensely competitive, dominated by well-funded giants like OpenAI, Google, and Anthropic, alongside strong Chinese contenders like DeepSeek, Alibaba, and Tencent. Xiaomi's 600-billion-yuan war chest is substantial, but sustained R&D at the frontier is notoriously costly. Its success will depend not only on continuous model innovation but also on its ability to attract a vibrant third-party developer community to its MiMo platform and seamlessly weave AI into its diverse product portfolio, realizing Lei Jun's vision of a deeply integrated "Human x Car x Home" intelligent ecosystem. The stealth launch of Hunter and Healer Alpha has successfully announced Xiaomi's arrival as a serious AI contender; the next three years will test its ability to stay there.