Xiaomi's large model claimed the 'global No. 1' spot—backed by capability, yet also aided by luck

(This article was written by Tang Chen and published by TMTPost with authorization)

By Tang Chen

Xiaomi has 'won' again—this time in large models.

Recently, Xu Jieyun, Xiaomi's head of public relations, reposted a ranking list. According to data from OpenRouter, a multi-model aggregation platform, the top five globally by API call volume last week (July 20–26) were all Chinese AI large models: Xiaomi MiMo-V2.5, DeepSeek V4 Flash, Tencent HY3, Zhipu GLM-5.2, and DeepSeek V4 Pro.

Xiaomi’s MiMo-V2.5, ranked first, recorded a weekly call volume of 10.5 trillion tokens, up 12% week-over-week. This model series entered public beta on April 23 and supports context windows at the million-token level, with multimodal capabilities spanning text, images, video, and audio.

It should be noted that this ranking has limitations: it excludes traffic from domestic internet giants’ proprietary apps, private enterprise deployments in China, and closed-API usage from major U.S. firms. Thus, it does not fully represent total global AI call volume but serves only as an observational window into the overseas open-source segment.

However, unlike internal corporate usage data, it aggregates real paid API calls from tens of thousands of independent overseas developers and small-to-midsize international AI applications, with all token usage measured under a unified standard—offering an objective reflection of actual developer preferences worldwide.

This also means that domestically developed open-source models are breaking through ecosystem barriers erected by foreign vendors and achieving leadership in fully commercialized third-party channels.

I’ve noticed that most mainstream media outlets are reportingthat Chinese-developed large models have consecutively held the top global market share for multiple weeks.OpenRouter disclosed that as of July 26, over the past 28 days, Chinese AI models accounted for 63.5% of global usage, compared to just 35.5% for U.S. models—the gap continuing to widen, with Chinese large models capturing nearly two-thirds of the global market.

Xu Jieyun’s official title is Special Assistant to the Chairman of Xiaomi Group and Deputy General Manager of the Strategic Marketing Department.He singled out MiMo-V2.5’s weekly achievement of being 'number one globally'—a move that, beyond showcasing Xiaomi's AI prowess, primarily reflects Xiaomi’s need for an AI-driven narrative to offset market sentiment stemming from its business challenges.

Xiaomi’s latest first-quarter 2026 earnings report officially stated, 'Operational quality continues to improve, and AI is accelerating the restructuring of our full ecosystem spanning people, vehicles, and homes.' However, the numbers show that both revenue and profit for the quarter fell short of market expectations.

For example, total revenue came in at RMB 99.1 billion, down 10.9% year-over-year; adjusted net profit was RMB 6.072 billion, a sharp 43.1% decline year-over-year. The day after the earnings release, Xiaomi’s stock experienced significant volatility—plunging sharply, rebounding quickly, then dropping again rapidly.

At this critical juncture where traditional businesses face mounting pressure on growth, MiMo’s claim to the 'global number one' title carries value that goes beyond mere rankings.Xiaomi needs to send a clear signal to the outside world: this is a milestone indicating that Xiaomi’s AI has moved beyond the pure investment phase and is now capable of large-scale commercialization and self-sustaining revenue generation.

Previously, AI had long functioned as a cost center with massive expenditures. The market has consistently questioned whether Xiaomi’s repeated emphasis on sustained AI investment was merely hype or represented genuine accumulation of substantive technological assets.

Xiaomi previously stated it would invest RMB 16 billion in AI in 2026, with cumulative investments in the field over the next three years amounting to no less than RMB 60 billion. This scale of investment is modest compared to leading native AI firms like OpenAI and Anthropic, as well as domestic giants such as Huawei, ByteDance, and Alibaba.

For instance, OpenAI projects it will spend approximately USD 750 billion on computing resources by 2030, exceeding its earlier 2026 projection of USD 600 billion. Meta has raised its full-year 2026 capital expenditure (CAPEX) guidance to between USD 125 billion and USD 145 billion, primarily for AI computing infrastructure.

Domestically, ByteDance has increased its 2026 capital expenditure budget for AI infrastructure to over RMB 200 billion—an increase of at least 25% from the internal baseline of approximately RMB 160 billion set at the end of 2025.

Alibaba is also highly aggressive, with projected full-year 2026 capital expenditures of RMB 150 billion, including approximately RMB 72 billion specifically allocated to computing power, focused on building out its Tongyi large model clusters and intelligent computing centers.

For Xiaomi itself, committing RMB 60 billion over three years already represents the optimal solution under resource constraints, as it runs multiple capital-intensive R&D initiatives in parallel.

Lei Jun and Xiaomi continue to reiterate this goal primarily to bolster confidence among the market and industry players. At the same time, he also hopes to solidify Xiaomi's identity as a hard-tech company.

Money must be spent, slogans must be shouted—and Xiaomi AI’s foundation is deeply imbued with 'Xiaomi characteristics.'

At his fourth annual speech in 2023, Lei Jun articulated the core logic of Xiaomi’s technology strategy:

(Software × Hardware)^AI。

This means deep investment in foundational technologies with sustained long-term commitment, deep integration of software and hardware, and comprehensive AI empowerment.

Under this framework, AI is not an independent commercial project for Xiaomi; rather, it serves as a multiplier tool designed to amplify the value of its software-hardware ecosystem. All AI achievements must be implemented within tangible hardware scenarios—smartphones, automobiles, and IoT devices.

At the execution level, Lei Jun’s vision has been concretely translated into the operational principles followed by MiMo business head Luo Fuli. She once stated that computing power and data will not constitute long-term barriers in AI; instead, the engineering capability to translate technology into physical products and convert it into commercial value is the truly irreplicable core competency.

This top-down technical doctrine has set MiMo on a 'pragmatic' engineering path.MiMo avoids reinventing the wheel on general-purpose large models, instead building upon mature open-source foundations. All its core in-house R&D resources are focused squarely on its competitive strengths: on-device hardware integration, cross-device compute orchestration, inference engineering optimization, and global commercialization.

This pragmatic strategic trade-off has enabled Xiaomi to sidestep the parameter arms race in large models and avoid direct competition for user traffic in generic conversational consumer products. All of its large model capabilities are embedded directly into its own hardware ecosystems for real-world deployment.

Across the industry, there is virtually no other company capable of simultaneously integrating a cloud-based commercial ecosystem with a closed-loop, billion-scale on-device hardware footprint.

Xiaomi sits precisely at the intersection of these two domains: pure-play AI ventures possess model development and API monetization capabilities but entirely lack physical hardware scenarios for deployment; meanwhile, other smartphone manufacturers control vast fleets of end-user devices yet continue to confine their AI capabilities strictly within walled-garden, on-device ecosystems, having not yet taken the critical step toward opening up global paid markets.

Specifically, leveraging Xiaomi’s HyperMind framework for pervasive compute orchestration within its HyperOS operating system, combined with hardware integration powered by its in-house Xuanjie chips, Xiaomi has built the world’s largest consumer-grade, offline, on-device AI cluster.

Hundreds of millions of terminal devices continuously generate real-world usage scenarios and user interaction data, which in turn drive iterative model improvements—establishing a virtuous closed loop of 'deployment → feedback → optimization' that effectively mitigates inherent limitations of open-source foundational models.

On the commercialization front, MiMo’s breakthrough is equally pragmatic. It deliberately avoids head-on competition with top-tier international closed-source models and instead precisely targets the core pain points of small and medium-sized developers both domestically and overseas. Through architectural optimizations that drastically reduce inference costs, it captures the mass market with high cost-performance, high-concurrency stability, and million-token ultra-long context capabilities—propelling it to the top of global benchmarks.

However, this 'world No. 1' achievement carries clear strategic limitations. MiMo’s edge lies in engineering deployment and commercial cost-efficiency—not in fundamental technological breakthroughs. Its capabilities in complex logical reasoning and quantitative problem-solving still lag behind those of leading international closed-source models.

Moreover, its low-margin, high-volume pricing strategy generates substantial scale but thin profitability, insufficient in the short term to recoup its tens-of-billions-of-yuan R&D investment. If Xiaomi continues prioritizing application-layer optimizations over the long term, it will remain in a follower position in foundational AI research.

In short,MiMo’s market breakthrough reflects both genuine capability and an element of good fortune.

A typical example is that Tencent’s large AI model, HunYuan H3, topped OpenRouter’s global large model ranking by call volume in its first week of launch. The key reason was that the model’s capabilities were solid—and it was offered free of charge for a limited time.

After unveiling his new 'human-vehicle-home' narrative, Lei Jun led Xiaomi to complete its hardware-software ecosystem and enhance interoperability among devices.

However, this ecosystem has always had a critical shortcoming: it offered device connectivity but lacked contextual intelligence. His solution has been to make long-term investments in three core technology areas—AI, chips, and operating systems (OS).

Among these, HyperOS addresses the systemic coordination challenges of multi-device collaboration, in-house chips fill the gap in on-device computing power, and AI transforms a cluster of cold hardware into an intelligent system capable of understanding users, anticipating needs, and autonomously coordinating actions.

Xiaomi’s 'technology iron triangle' has thus been fully assembled: the OS serves as the skeleton, chips as the muscles, and AI as the brain of the ecosystem—completing the final piece of the human-vehicle-home puzzle.

At a time when hardware differentiation is minimal, premium market penetration remains challenging, and its automotive business is still scaling up, AI’s strategic value is becoming increasingly evident.

To some extent, MiMo’s achievement of securing the top spot in global paid API calls represents a validation of Xiaomi’s pragmatic AI strategy.

Xu Jieyun may still be laying the groundwork for what’s being called 'Xiaomi’s full-stack in-house integration.' Xiaomi Group President Lu Weibing revealed thatXiaomi’s self-developed chips, operating system, and large AI models will achieve deep integration within this year and unleash their combined capabilities on a single terminal product.

Industry observers believe this move signals that Xiaomi is building a full-stack, end-to-end technology architecture—from chips to operating system to AI—forming a complete closed-loop capability.

This first-place achievement by Xiaomi Mimo may well be the prelude to this story.

References:Jiemian News, 'The Top Five Most Used Global Large Models Are All Chinese Products'

79K Views