Xiaomi AI
Last reviewed
Sources
36 citations
Review status
Source-backed
Revision
v4 · 4,137 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Sources
36 citations
Review status
Source-backed
Revision
v4 · 4,137 words
Add missing citations, update stale details, or suggest a clearer explanation.
Xiaomi AI is the artificial intelligence research, products, and strategy of Xiaomi Corporation, the Beijing-based consumer-electronics and electric-vehicle maker, spanning open-source large language models (the MiMo and MiLM families), the XiaoAI voice assistant, the HyperOS operating system and its on-device HyperAI features, autonomous-driving software for the Xiaomi SU7 and YU7 electric vehicles, and the CyberOne and CyberDog robots. Xiaomi's headline AI result is MiMo-7B, an open-source reasoning model released in April 2025 that, with only 7 billion parameters, matches OpenAI's o1-mini and beats Alibaba's 32-billion-parameter QwQ-Preview on mathematics and coding benchmarks. [1][7][31] Xiaomi was founded on April 6, 2010, by Lei Jun and six co-founders, and by late 2024 connected over 900 million smart devices to its AIoT platform, placing the company at the intersection of consumer hardware and AI infrastructure. [1][3]
Xiaomi was co-founded by Lei Jun alongside Lin Bin (former vice president of Google China's Institute of Engineering), Zhou Guangping (former senior director at Motorola's Beijing R&D center), Liu De (former chair of Industrial Design at the University of Science and Technology Beijing), Li Wanqiang (former general manager of Kingsoft Dictionary), Huang Jiangji (principal development manager), and Hong Feng (former senior product manager at Google China). Lei Jun, who had previously served as CEO of Kingsoft Corporation and led it through its 2007 Hong Kong Stock Exchange IPO, envisioned a company that could deliver high-quality smartphones at affordable prices. [1][2]
Xiaomi released its first Android-based firmware, MIUI, on August 16, 2010. The company's first smartphone, the Mi 1, followed in 2011. Over the following decade, Xiaomi expanded into tablets, laptops, smart TVs, wearables, home appliances, and a wide variety of IoT products. The company went public on the Hong Kong Stock Exchange in July 2018. By 2024, Xiaomi's annual revenue had grown substantially, and its product portfolio included electric vehicles, with the Xiaomi SU7 sedan launching in March 2024. [1][18]
Xiaomi established its AI Lab in 2017, starting with a team of just four researchers. The lab's initial work focused on speech recognition, and it quickly delivered Xiaomi's first speech recognition system for use in smart TV voice interactions. Over the following years, the AI Lab expanded its research into computer vision, natural language processing, acoustics, knowledge graphs, machine learning, and multimodal systems. [3]
By 2023, Xiaomi's AI-related teams had grown to over 3,000 members. In April 2023, the company formed a dedicated Foundation Model team within the AI Lab, consisting of over 1,200 employees focused on developing large language models. In mid-November 2024, Xiaomi's Basic Technology Platform Department created a separate AI Platform Department, led by Duo Zhang, to coordinate the company's expanding AI efforts across all product lines. [3]
Xiaomi has also invested heavily in recruiting top AI talent. In early 2025, reports indicated that Fuli Luo, one of the key developers behind DeepSeek-V2, was expected to join Xiaomi's AI research division, signaling the company's intent to compete at the frontier of large language model development. [28]
Xiaomi's first publicly disclosed large language model, MiLM-6B, appeared on benchmark leaderboards in August 2023. The model has 6.4 billion parameters and was designed as a lightweight model suitable for deployment on mobile devices. On the CMMLU (Chinese Massive Multitask Language Understanding) benchmark, MiLM-6B ranked first among Chinese large models. On the C-Eval benchmark, MiLM-6B placed first within its parameter category and ranked tenth overall across all model sizes. The model showed particular strength in STEM fields, achieving high accuracy across 20 science and engineering subjects. [4]
In May 2024, Xiaomi announced that MiLM had officially passed China's large model filing requirements, clearing the way for integration across the company's smartphones, automobiles, and smart home products. [5]
The MiLM2 series, released in November 2024, continued Xiaomi's emphasis on lightweight, deployable models. MiLM2 supports a range of parameter scales from 0.3 billion (suitable for on-device use) to 30 billion (for cloud-based tasks). Xiaomi reported that MiLM2 achieved a 45% average improvement over its predecessor across ten capabilities, including instruction following, translation, and conversational ability. [6]
In April 2025, Xiaomi released MiMo-7B, its first open-source reasoning-focused language model, under the tagline "Unlocking the Reasoning Potential of Language Model." [31] Despite its relatively small size of 7 billion parameters, MiMo-7B delivered strong results on mathematical reasoning and code generation benchmarks. The reinforcement-learned MiMo-7B-RL variant scored 95.8% (Pass@1) on MATH-500, 68.2% on AIME 2024, 55.4% on AIME 2025, and 57.8% on LiveCodeBench v5. [31] According to Xiaomi's model card, MiMo-7B-RL "demonstrates superior performance on both mathematics and code reasoning tasks, matching the performance of OpenAI o1-mini" while surpassing Alibaba's much larger 32-billion-parameter QwQ-32B-Preview. [31][7] Training involved approximately 25 trillion tokens across three phases, including a specialized dataset of 200 billion reasoning tokens. Xiaomi released four checkpoints (Base, SFT, RL-Zero, and RL) on Hugging Face under the Apache 2.0 license. [7][31]
In June 2025, Xiaomi followed with MiMo-VL-7B, a multimodal vision-language model that pairs a native-resolution Vision Transformer encoder with an MLP projector and was trained with a four-stage pre-training run of about 2.4 trillion tokens plus Mixed On-policy Reinforcement Learning (MORL). Xiaomi released the checkpoints and a 50+ task evaluation suite on GitHub. [32] A specialized home-centric variant, MiMo-VL-Miloco-7B, targets smart-home video and scene understanding. [32]
In December 2025, Xiaomi followed up with MiMo-V2-Flash, a significantly larger Mixture-of-Experts model with 309 billion total parameters and 15 billion active parameters. Released under the MIT open-source license, MiMo-V2-Flash uses a hybrid attention architecture that interleaves Sliding Window Attention and Global Attention in a 5:1 ratio with a 128-token window, reducing key-value cache storage by nearly 6x while maintaining strong long-context performance. [8][33] On the SWE-Bench Verified benchmark for software engineering, the model scored 73.4%, ranking first among open-source models. [33] On the AIME 2025 mathematical reasoning benchmark, it achieved 94.1%, close to proprietary models such as GPT-5 High (94.6%). [8] A multi-token prediction head lets the model reach inference speeds of around 150 tokens per second, roughly 2 to 3 times faster than comparable models. [33]
The following table summarizes Xiaomi's major language models:
| Model | Release | Parameters | Key Benchmarks | License |
|---|---|---|---|---|
| MiLM-6B | August 2023 | 6.4B | CMMLU 1st (Chinese LMs), C-Eval 1st (same size) | Proprietary |
| MiLM2 | November 2024 | 0.3B to 30B | 45% avg. improvement over MiLM | Proprietary |
| MiMo-7B | April 2025 | 7B | MATH-500: 95.8%, AIME 2024: 68.2%, LiveCodeBench v5: 57.8% | Apache 2.0 |
| MiMo-VL-7B | June 2025 | 7B | Multimodal (vision-language); 50+ task suite | Open Source |
| MiMo-V2-Flash | December 2025 | 309B (15B active) | SWE-Bench Verified: 73.4%, AIME 2025: 94.1% | MIT |
To support its growing AI ambitions, Xiaomi announced in late 2024 that it was constructing a dedicated GPU cluster comprising 10,000 GPUs for training large AI models. Prior to this expansion, the company's Foundation Model team had access to approximately 6,500 GPUs. Xiaomi CEO Lei Jun has been directly involved in overseeing this infrastructure buildout. [9]
The investment in computing infrastructure reflects a broader strategic commitment. Xiaomi spent approximately RMB 24 billion (about $3.3 billion) on R&D in 2024 and announced plans to increase that figure to RMB 30 to 33 billion ($4.1 to $4.6 billion) in 2025, with a projected RMB 40 billion ($5.6 billion) in 2026. [10][11] Lei Jun has said that AI and related business would account for roughly a quarter of Xiaomi's planned RMB 30 billion R&D budget in 2025. [34] Over five years, Xiaomi plans to invest RMB 200 billion ($27.8 billion) in research and development. As of September 2025, the company employed over 24,000 R&D personnel. [11]
Xiaomi replaced its long-running MIUI operating system with HyperOS, which launched alongside the Xiaomi 14 series in October 2023. HyperOS serves as the unified software platform across Xiaomi's smartphones, tablets, wearables, smart home products, and vehicles, forming the backbone of the company's "Human x Car x Home" ecosystem strategy. [12]
HyperOS 2.0 debuted on October 29, 2024, alongside the Xiaomi 15 series and the SU7 Ultra electric vehicle. Built on Android 15, it introduced a suite of on-device AI features branded as "HyperAI." Xiaomi formally introduced HyperAI to global markets at Mobile World Congress (MWC) 2025 in Barcelona. [12][13]
HyperAI features include:
HyperOS 3 launched in China on August 28, 2025, with a global release following on September 24, 2025. It deepened the integration of on-device AI, including a Deep Think Mode for extended reasoning, background noise reduction, instant recording summaries, and AI-powered search that can summarize files and scan locally stored media. The upgraded XiaoAI assistant in HyperOS 3 supports over 80 apps and can perform more than 1,000 actions, backed by both on-device and cloud-based large language models. [14]
XiaoAI (also written Xiao AI) is Xiaomi's voice assistant, first introduced in September 2017 alongside the Xiaomi TV 4A. By August 2021, XiaoAI had surpassed 100 million monthly active users. [15] With HyperOS 2, Xiaomi launched "Super XiaoAI," a next-generation version of the assistant powered by the company's own large language models. Super XiaoAI can handle on-screen content recognition, natural language commands, and contextual follow-up conversations. Users can circle content on their screen to trigger instant actions such as translation and web search.
Xiaomi operates one of the world's largest consumer IoT platforms. As of the end of 2024, approximately 904.6 million smart devices (excluding smartphones, tablets, and laptops) were connected to Xiaomi's AIoT platform, reflecting 22% year-over-year growth. By the third quarter of 2025, this number exceeded one billion devices, with over 100 million users having five or more connected Xiaomi products. [30]
The ecosystem spans over 200 product categories and covers roughly 95% of daily life scenarios. Products include smart speakers, routers, TVs, air conditioners, refrigerators, washing machines, rice cookers, robot vacuum cleaners, air purifiers, lighting systems, smart sockets, and motorized curtains. The Mi Home app serves as the central control interface for all connected devices.
Xiaomi Vela is an IoT software platform built on top of Apache NuttX, a real-time operating system (RTOS) maintained under the Apache Software Foundation. Xiaomi introduced Vela at its annual MIDC developer conference in 2020. Because NuttX is an extremely lightweight RTOS designed for resource-constrained embedded devices, Vela can run on hardware ranging from simple sensors to wearables and small appliances. [17]
As of 2025, Xiaomi Vela has been integrated into over 140 million devices worldwide and supports more than 354 chip platforms. Xiaomi open-sourced Vela as the OpenVela project under the Apache 2.0 license, with code available on both GitHub and Gitee. Xiaomi developers have contributed over 50% of community code contributions for three consecutive years. [16]
At MWC Barcelona 2024, Xiaomi unveiled its "Human x Car x Home" platform, a strategic framework that positions the smartphone as the gateway to a fully connected lifestyle linking personal devices, smart home products, and vehicles. HyperOS serves as the operating system foundation for this ecosystem, enabling real-time cross-device networking and unified control through a shared codebase. [12]
Xiaomi HyperConnect is the connectivity layer that enables these cross-device interactions, supporting features such as dual-camera streaming between devices, "Home Screen+" for extending displays across form factors, and seamless handoff between smartphones, tablets, vehicles, and home appliances.
The Xiaomi SU7 is a full-size four-door fastback electric sedan and the first motor vehicle developed by Xiaomi. It was unveiled on December 28, 2023, and officially launched on March 28, 2024, with deliveries beginning on April 3, 2024. [18] The SU7 received 88,898 orders in its first 24 hours. [18] The SU7 represents Xiaomi's entry into the autonomous driving sector and serves as a key platform for the company's automotive AI technologies.
The SU7's higher-end variants feature Xiaomi HAD (Highways Autonomous Driving), which provides supervised autonomous driving including highway Navigate on Autopilot (NOA), hands-free parking, and car summoning. Subsequent over-the-air (OTA) updates expanded NOA capabilities to urban driving conditions.
Xiaomi has developed several proprietary AI systems for the SU7:
| Technology | Description |
|---|---|
| End-to-End AI Parking | The world's first production-ready end-to-end sensing and decision-making AI model for automated parking, capable of navigating parking garages with elevators |
| Road-Mapping Foundation Model | Real-time road understanding that allows the vehicle to navigate complex intersections without relying on pre-built HD maps |
| Super Res Occupancy Network | Machine learning-based real-time occupancy analysis for safe decision-making around obstacles and other road users |
The original SU7 Max is equipped with two NVIDIA Drive Orin chips providing a combined 508 TOPS of computing power, one LiDAR sensor, eleven high-definition cameras, three millimeter-wave radars, and twelve ultrasonic radars. [18][19] The Xiaomi Pilot Pro configuration uses a single Drive Orin chip with 84 TOPS, while Pilot Max uses two for the combined 508 TOPS. [18] The LiDAR system is supplied by Hesai Technology. [19]
The next-generation SU7, announced in March 2026, makes LiDAR standard across all variants and upgrades to a unified computing platform delivering 700 TOPS of processing power. It also adds 4D millimeter-wave radar and achieves a range of up to 902 km on a single charge. [20]
The Xiaomi SU7 quickly became one of the best-selling new EV entrants in China. In 2024, Xiaomi delivered approximately 139,547 SU7 units, exceeding its production target of 130,000 units (which had already been raised twice from initial goals of 76,000 and then 120,000). [29] For 2025, Xiaomi set a delivery target of 300,000 vehicles and ultimately delivered over 400,000 units across the SU7 and the newer YU7 SUV. [21] The YU7, launched on June 26, 2025, drew roughly 200,000 firm orders within its first three minutes on sale. [35]
Xiaomi CyberDog is a quadruped robot unveiled in August 2021. Described as a "bio-inspired quadruped robot" designed for the open-source community and developers worldwide, CyberDog serves as both a research platform and a robotic companion. [22]
| Specification | Details |
|---|---|
| Weight | Approximately 3 kg |
| Walking Speed | 5.8 km/h |
| Running Speed | 11.5 km/h (approximately 3.2 m/s) |
| Processor | NVIDIA Jetson Xavier (6-core Carmel ARM v8.2 CPU, 384-core Volta GPU with 48 Tensor Cores) |
| Memory | 8 GB 128-bit LPDDR4x |
| Storage | 16 GB eMMC + 128 GB SSD |
| Sensors | 11 high-precision sensors including Intel RealSense D450 depth camera, AI cameras, binocular ultra-wide cameras, touch sensors, GPS |
| Capabilities | Autonomous navigation, object tracking, SLAM, obstacle avoidance, face and voice recognition |
| Control Methods | Mobile app, voice control, remote control |
Xiaomi released CyberDog 2 on August 14, 2023. The second generation features a complete design overhaul inspired by the Doberman breed, with a body measuring 562 x 339 x 481 mm and weighing 8.9 kg. It is powered by the NVIDIA Jetson Xavier NX, which includes 384 CUDA cores, 48 Tensor Cores, and a 6-core Carmel ARM CPU, paired with 8 GB of LPDDR4x RAM and 32 GB of SSD storage. [23]
CyberDog 2 carries 19 high-precision sensors, including an RGB camera, an AI camera, four Time-of-Flight (ToF) sensors, a LiDAR sensor, a depth camera, and an ultrasonic sensor. It uses Xiaomi's custom-developed CyberGear micro-actuators for improved agility, enabling maneuvers such as continuous backflips and self-recovery from falls. The robot underwent more than 30,000 AI simulations of real dog behavior to achieve natural, lifelike movement. Xiaomi open-sourced CyberDog 2's code, structural drawings, and modular sensor processing capabilities. [23]
CyberOne is Xiaomi's humanoid robot, unveiled at a product launch event in Beijing on August 11, 2022, by Lei Jun himself, who received a flower from the robot on stage. [24][25] All of CyberOne's AI and mechanical capabilities were developed in-house by the Xiaomi Robotics Lab.
| Specification | Details |
|---|---|
| Height | 177 cm |
| Weight | 52 kg |
| Arm Span | 168 cm |
| Degrees of Freedom | 21 |
| Payload (single hand) | Up to 1.5 kg |
| Upper Limb Actuator Weight | 500 g each, rated output torque up to 30 Nm |
| Processor | NVIDIA Jetson Xavier NX |
| Sensors | 11 built-in sensors, Mi-Sense depth vision module |
| Emotion Recognition | 45 classifications of human emotion |
| Sound Recognition | 85 types of environmental sounds |
| Battery Life | 2 to 3 hours active use, 4 to 5 hours standby; 90-minute charge time |
| Capabilities | 3D spatial perception, gesture recognition, expression detection, walking and balancing on rough terrain |
CyberOne uses a proprietary Mi-Sense depth vision module combined with a MiAI environment-semantics engine and a MiAI vocal-emotion engine to perceive three-dimensional space and recognize people, gestures, 85 types of environmental sound, and 45 classifications of human emotion. [24][36] The robot represents Xiaomi's long-term bet on embodied AI and general-purpose humanoid platforms for industrial and research applications.
Xiaomi's semiconductor efforts date back to 2014, when it established Pinecone Electronics as a chip design subsidiary. The company released its first mobile system-on-chip, the Surge S1, in 2017. Subsequent specialized chips include the Surge C1 image signal processor (ISP), launched in March 2021 for the MIX FOLD foldable phone, and the Surge P1 fast-charging controller, introduced in December 2021. [26]
In 2025, Xiaomi unveiled the Surge Xuanjie O1, a self-developed 3nm SoC marking the company's return to full application processor development after an eight-year gap. As of April 2025, Xiaomi had invested RMB 13.5 billion (approximately $1.87 billion) into the Xuanjie project, with an R&D team exceeding 2,500 engineers. The company has committed to investing at least RMB 50 billion (roughly $7 billion) over a decade into chip development. [27]
While Xiaomi's chip efforts have so far focused on application processors, ISPs, and power management, the company's expanding AI infrastructure and on-device model deployment suggest that dedicated AI acceleration hardware could become a future priority.
Xiaomi competes with Samsung, Apple, Google, and other major consumer electronics companies in the race to bring AI capabilities to everyday devices. Each company has taken a different approach:
| Company | AI Suite | Key Features | On-Device Focus |
|---|---|---|---|
| Samsung | Galaxy AI | Real-time call translation, AI photo editing (Generative Fill), note summarization | Hybrid (on-device + cloud) |
| Apple | Apple Intelligence | Writing tools, Siri with ChatGPT integration, generative image tools | Strong on-device emphasis |
| Xiaomi | HyperAI | AI writing, real-time translation, Magic Eraser, AI video creation, XiaoAI assistant | Hybrid (on-device + cloud) |
| Gemini | Multimodal understanding, Circle to Search, call screening, Pixel-exclusive features | Hybrid (on-device + cloud) |
Xiaomi differentiates itself through the sheer breadth of its ecosystem. While Samsung and Apple primarily integrate AI into phones, tablets, and wearables, Xiaomi's AI capabilities extend across smartphones, smart home devices, vehicles, and robots, all connected through HyperOS. The company's willingness to open-source its models (MiMo-7B, MiMo-V2-Flash) and IoT platform (OpenVela) also distinguishes it from competitors with more closed ecosystems.
In the Chinese market specifically, Xiaomi competes with Baidu, Huawei, and other domestic technology firms that are also investing heavily in large language models and on-device AI. Xiaomi's advantage lies in its massive installed base of IoT devices, which provides both a distribution channel for AI features and a source of real-world usage data.
Xiaomi's AI strategy is built around the concept of turning its hardware ecosystem into an AI-powered platform. The company has articulated several key strategic priorities:
On-device AI deployment: Xiaomi prioritizes running AI models locally on devices for faster response times and stronger data privacy. The MiLM2 series and MiMo models are specifically optimized for efficient inference on resource-constrained hardware.
AI across all product lines: Through the Human x Car x Home framework, Xiaomi aims to embed AI in every product category it touches, from phones and home appliances to electric vehicles and robots.
Open-source engagement: By releasing MiMo models and the Vela IoT platform as open-source projects, Xiaomi seeks to build developer communities and accelerate ecosystem growth.
Autonomous driving: The SU7's AI driving capabilities represent a major investment area, with Xiaomi developing end-to-end models for parking, navigation, and occupancy recognition.
Embodied AI and robotics: CyberDog and CyberOne serve as research platforms for exploring physical intelligence, with potential long-term applications in service, industrial, and consumer robotics.
Xiaomi has framed AI, operating systems, and chips as the three fundamental technologies behind a planned investment of more than RMB 100 billion between 2022 and 2026. [34] With planned R&D spending of $5.6 billion in 2026 and a five-year commitment of $27.8 billion, Xiaomi is positioning itself as a serious contender in the global AI landscape, not just as a smartphone maker with AI features, but as a full-stack AI company with capabilities spanning models, chips, operating systems, and hardware. [11]