Principal Activities

The principal activities of the Group are the sale of advanced AI software platform and related services, sale of AI software-embedded hardware and related services, AIDC service as well as research and development activities in relation to AI technology.

Latest Results

The Group's loss attributable to shareholders for the year ended 31-12-2025 amounted to RMB 1.77 billion. Basic loss per share was RMB 0.0476. No dividend declared. Turnover amounted to RMB 5.01 billion, an increase of 32.9% over the same period last year, gross profit margin down 1.9% to 41.0%. (Announcement Date: 24 Mar 2026)

Business Review - For the year ended December 31, 2025

Generative AI

Large Model R&D: We have established a leading position with cutting-edge technologies in multi-modal large model for our proprietary architecture

In 2025, SenseTime continued to invest in R&D of cutting-edge technologies, comprehensively establishing its leading position in the domestic multimodal large model sector. The SenseNova series of large models continuously raised the ceiling for multimodal reasoning, with the SenseNova-SI Spatial Intelligence Model being the first to break down boundaries between the virtual and the real, ultimately culminating in the release of the new NEO architecture, which fundamentally reshaped the underlying logic of native multimodality. Collectively, these breakthroughs not only redefined model inference efficiency and training paradigms, but also marked SenseTime’s entry into a new era in its pursuit of multimodal integration.

1. SenseNova Multimodal Large Model Series

In April 2025, the SenseNova V6 became China’s first native multimodal general-purpose model integrating language and vision, topping both SuperCLUE (language) and OpenCompass (multimodal) benchmarks with one single model.

SenseNova V6.5 launched in July 2025, achieving early fusion at the encoder level, reducing the visual encoder to 1B parameters and improving cost-performance by three-fold.Meanwhile, it spearheaded commercial-grade interleaved text-image reasoning in China and integrated imagistic thinking into the large model reasoning process. SenseNova V6.5 secured its leading position in China in SuperCLUE Benchmark in December 2025, and achieved the highest score nationwide in visual reasoning, and took the lead in seven sub-tasks including scientific reasoning, logical reasoning and code design.

Building on such foundation, we further launched an agent that combines visual reasoning with deep search, and open-sourced SenseNova-MARS in Jan 2026. It outperformed Gemini-3-Pro and GPT-5.2 in multimodal searching and reasoning benchmarks, becoming the first Agentic VLM model to support deep fusion of dynamic visual reasoning and image/text search. Such achievement represents a leading international level in the field of autonomous intelligent agents and has already begun commercial deployment in city-level application scenarios.

2. SenseNova-SI Spatial Intelligence Model Series

Building on our long-standing expertise in Computer Vision, we systematically built a spatial intelligence framework, which seeks to break down boundaries between the digital world and physical world, empowering application scenarios such as embodied intelligence and autonomous driving. In Nov 2025, we released and open-sourced the SenseNova-SI Spatial Intelligence Model. In a comprehensive evaluation integrating several internationally authoritative spatial intelligence metrics, SenseNova-SI-1.3 demonstrated outstanding performance and ranked first globally among peer models (as of Feb 2026). Its overall performance outperformed Gemini-3-Pro and Cambrican-S across multiple authoritative benchmark tests. It also tops the ranking in HuggingFace in terms of download volume. In December, we jointly released and open- sourced the Kaiwu world model Kairos-SenseNova with ecosystem enterprises, becoming the first open-source embodied native world model to achieve integrated“multimodal understanding–generation–prediction”, in the globally recognised Embodied Intelligence Benchmark, Kairos 3.0-4B outperformed all other models across the board. In the A800 GPU Benchmark, thanks to its model capabilities and inference tools, Kairos 3.0-4B achieved an inference speed seventy-two times faster than Cosmos 2.5, setting a new global performance record for embodied world models and fully demonstrating its cutting-edge technical prowess.

3. Launched the Integrated Multimodal Model Architecture–NEO

In December 2025, SenseTime unveiled and open-sourced NEO, an entirely new native multimodal model architecture. As the industry’s first commercially-viable native multimodal architecture (Native VLM), NEO evolved beyond the mainstream“encoder-connector-LLM hub”splicing architecture. Designed from first principles, NEO features three core innovations: Native Embedding, Native RoPE, Native Multi-head Attention, to achieve the fusion of textual and visual information at the core level. NEO also delivers extraordinary learning efficiency, it achieves state-of-the-art performance comparable to peer models of similar scale in the industry with only one-tenth the training data and computing power.

We will launch a new model based on the second-generation NEO architecture in the second quarter of 2026, taking the lead in the industry in validating a new“Scaling Law”under the native multimodal architecture that deeply unifies understanding and generation. We expect this model to play a significant role in areas such as full-modality reasoning, perception and generative interactive creation, and spatial intelligence. Leveraging its powerful visual reasoning capabilities, it will significantly enhance the ability of agents to process multimodal information including images, videos, documents and webpages, enabling agents to handle complex scenario tasks more efficiently and deeply empowering agent AI and embodied intelligence applications.

Large Model Applications: Penetrating High-Value Scenarios and Securing Strategic Positions in Vertical Applications

By deeply understanding industry demands and continuously upgrading product capabilities, we have established leadership positions across multiple high-value large-model application segments.Thanks to our outstanding market performance, SenseTime’s Little Raccoon, Kapi Camera, and Kapi Accounting have been prominently featured on the“AI Efficiency”and“AI Entertainment/ Application”lists of the authoritative tech media QuantumBit’s“2025 Flagship AI 100”rankings.

1. High-Value Productivity Scenarios

We leverage a comprehensive framework combining task complexity and fault tolerance to unlock the closed-loop value of agents across strategic scenarios such as office, education, finance, marketing and content generation:

•In general office scenarios, SenseTime Raccoon has served over 15 million individual users and thousands of enterprise customers. In 2025, the Raccoon series recorded a seven-fold increase in monthly active users (MAUs), fully validating strong market appeal and acceptance. In product evolution, we have realized processing of million-scale data in seconds and autonomous generation of high-quality, editable PPT documents. Our enterprise-grade data analytics solution delivers 95% analysis accuracy and reduces analysis duration by 90%, greatly improving office efficiency.

•In the e-commerce marketing sector, we now cover the entire e-commerce workflow.AI is no longer a mere“chat bot”, but becomes the“hands”to manage backends and execute tasks, enabling end-to-end autonomous implementation from decision-making to execution. For example, we have formed an in-depth strategic cooperation with SMZDM, helping it improve store operation efficiency by twenty times, live broadcast operation efficiency by six times, traffic placement efficiency by five times, and real-time monitoring efficiency by three times.

•In content creation, SenseTime launched Seko in 2025, the industry’s first full-process short drama creation agent integrating scripting and production. It directly generates and creates single episode up to three minutes long and can continuously generate a hundred episodes. Seko quickly attracted over 300,000 creators and reached 100,000 monthly active users (MAUs) within just a few months from launch. The live-action short drama Wan Xin Ji, powered by Seko, topped the Douyin AI Short Drama Chart upon its release in October 2025, becoming a phenomenal hit.

•In the financial sector, centering on the“Financial Intelligent Decision-Making Platform”as our core delivery vehicle, and through a differentiated model of“standardized products + ecosystem-driven delivery”, we have successfully addressed the industry’s pain points of heavy reliance on labor and limited scalability, while achieving low-cost monetization by leveraging our domestic adaptation advantages.In the banking sector, we have successfully penetrated client scenarios through the“Frontier Deployment Engineer”(FDE) model, achieving rapid revenue expansion across institutions such as Ping An Bank and Bank of Ningbo. In the insurance sector, we entered the trillion-RMB claims market in 2025, using it as an entry point to advance closed-loop industry exploration across underwriting and pricing, marketing and services, as well as healthcare.

•In empowering intelligent terminals, we leverage the core strengths of the“SenseNova”large model in full-modality interaction, low latency and long memory to drive the deep integration of multimodal large models into various terminal hardware, serving as interaction nodes between AI and the physical world. At the same time, we have enabled“pan-terminal hardware”such as smartphones, AI glasses, in-vehicle interaction systems and embodied intelligence, with customers including Xiaomi, BYD and Transsion.

2. AI-Native Consumer APPs for Young Users

In 2025, SenseTime made progress in exploring a new generation of AI-native consumer applications, for example, the Kapi product series successfully accumulated users reaching the tens of millions scale, validating the explosive growth potential of AI-native applications.

Kapi Camera, the industry’s first AI camera assistant built on multimodal large models. As a globally launched original Chinese product, it has topped the free charts on App Stores in multiple countries worldwide. As an AI financial assistant for young people, Kapi Budgeting saw its user base grow rapidly in 2025. Its T+1 retention rate steady at 70%–an industry- leading level–making it a benchmark for the seamless integration of SenseTime’s AI technologies into daily life. Both products were selected for the QbitAI 2025 AI 100 Flagship Products List, among which Kapi Camera ranked among Quantum Bit’s Top 10 Newly Added Downloaded Apps in the“AI Entertainment & Lifestyle”category.

SenseCore AI Infrastructure: Dual Engines of Capability Building and Scenario-based Deployment Forge a Solid Foundation for AI Productivity

As the solid foundation of our“Infrastructure–Model–Application”three-in-one strategy, SenseCore achieved a profound transition from technological strengths to an industrial closed loop in 2025. Through in-depth coordination between underlying computing power and algorithm R&D, it supported nearly one million model R&D tasks throughout the year. In the China Full-Stack AI Cloud Service Market Report H1 2025 released by Frost & Sullivan and LeadLeo Research Institute, SenseCore ranked among the top 4 in China’s full-stack AI cloud service market and ranked first among native AI cloud vendors, continuing to lead the native AI cloud vendor sector.

1. Advancing Infrastructure Capability to Upgrade the Computing Power Foundation from “Available”to“Effective”

SenseCore has unlocked the full chain from underlying hardware to top-level applications and from software stacks to model adaptation, accelerating the formation of a positive cycle in the AI industry:

•In-depth hardware-software co-optimization for differentiated advantages: In capability building, the core strength of SenseCore lies in its high-level optimization for hardware and software for large model R&D. The in-depth linkage of“computing power, model and application”forms our most robust technological moat. While enabling continuous breakthroughs and generational leadership in model development, it provides performance assurances and efficiency for upper-layer applications. Taking the LightX2V World Model Inference System as an example, leveraging cutting-edge technologies including step distillation and extreme quantization, we took the lead in realizing real-time video generation for physical world simulation. Amid chip process constraints, we also outperformed overseas top chips on domestic hardware. With its exceptional computing efficiency, the system has been officially recommended by various leading developers such as Alibaba and Tencent. Its open-source model has been downloaded over ten million times on HuggingFace, consistently ranking among the top ten globally (on par with OpenAI gpt-oss).

•Domestic ecosystem collaboration and model closed loop: We partnered with over a dozen of chip suppliers including Huawei Ascend, Hygon and Cambricon to launch the“SenseCore Computing Power Mall”, enabling flexible combination and allocation of diverse computing resources, platform tools and industry model services. SenseCore took the lead in completing full adaptation to the Ascend 384 super-node, possessing end-to-end delivery capabilities from liquid-cooled clusters to AI platforms. Kaiwu Kairos-SenseNova, a landmark domestic world model, has realized full-process training on domestic computing clusters, becoming the first world model to generate dynamic scenes in real time on the domestic chip ecosystem.

•Global export of expertise: SenseCore has translated its leading domestic intelligent computing expertise into global competitiveness by launching China’s first overseas domestic computing cluster in Saudi Arabia. By exporting our highly adaptable software stack, this project has validated our exceptional delivery capabilities in heterogeneous environments and the commercial potential of our asset-light global expansion model.

Since 2025, SenseCore has secured multiple authoritative certifications in safety and compliance, platform capabilities and computing performance, further consolidating its industry-leading position. In February 2026, its native AI cloud platform was awarded the“Excellent Level”certification in the Software Supply Chain Security Capability Assessment by the Ministry of Industry and Information Technology (MIIT), becoming one of the first enterprises nationwide to receive this honor. In January 2026, in the authoritative test of“Computing-Model-Application Integration–Computing Platform Service Capability”jointly conducted by the China Academy of Information and Communications Technology (CAICT) and China Telecommunication Technology Labs (CTTL), the platform received the industry’s first“5A Excellence Level”certification. In the 2025 China Large Model Inference Computing Power Market Analysis Report released by IDC in August 2025, SenseCore achieved full marks in all core dimensions, including“performance optimization”,“cloud-on-premise integration”and“large model training and inference”.

2. Scenario-based Solutions Bridge the“Last Mile”of AI Commercialization

Leveraging our end-to-end capabilities in“computing power–platform–solutions–services”, we have built a strong industrial cluster advantage in key sectors such as cutting-edge industrial R&D, AI4S (AI for Science), internet, interactive video content generation, embodied intelligence, smart hardware, and industrial intelligence, as part of our implementation of the national“AI+”initiative. SenseCore has become a core partner for numerous leading domestic research institutions, internet technology giants, top-tier Pan- entertainment companies, embodied intelligence robotics firms, and first-tier large-model unicorns.

3. Significantly Reducing Energy Consumption Costs by Pioneering the“Computing Power Synergy Large Model×Energy Dispatch Engine”

As of the release of this result announcement, the total operational computing scale of SenseCore has reached 40,400 PetaFLOPS (FP16). While pursuing ultimate computing efficiency, we actively respond to the national call for“computing-power synergy”.With the support of CATL, we have developed the world’s first AI intelligent system to achieve full-chain integration of“computing power management–IDC operation–energy storage systems”. By leveraging large models to analyze computing task data, the system can accurately predict power load and dynamically generate the optimal energy dispatch strategy. We estimate that the system will deliver a 7% reduction in electricity costs and cut carbon dioxide emissions by over 4,000 tonnes. This practice not only sets a technological benchmark for green computing in the industry, but also provides a replicable and verifiable SenseTime solution for low-carbon operation of large-scale intelligent computing centers.

Computer Vision

After years of strategic business optimization, the Group’s Computer Vision has successfully broken through the boundaries of traditional“perceptual intelligence”and been fully upgraded into a spatial intelligence and visual decision-making hub–powered by multimodal large models as its core engine–that deeply understands and reshapes interactions with the physical world.This marks that SenseTime Computer Vision business is moving from the technology investment phase into a period of large-scale commercial returns, becoming a solid pillar driving the Group’s revenue growth and cash flow improvement.

•Global leadership: SenseTime has ranked first in China’s computer vision market share for 9 consecutive years. Recognized for its strong global footprint, the Company was named a“Flagship Brand”in the 2025 Forbes China Go-International Series Selection. Customers from Southeast Asia, Northeast Asia and Middle East continued to repurchase our products and services, while new customers from South America and Europe started to pay more attention to our products, validating the strong influence of our brand and the adaptability of our products across various markets.

•Upgrade of SenseFoundry Platform: Building on more than a decade of expertise and technical heritage, the Computer Vision business has unlocked its second growth curve by redefining its SenseFoundry Platform. We have established a new multimodal algorithm production system centered on“integrating general and specialty capabilities, and unifying skills training and multimodal large model”. With multimodal large model agents as the core, it integrates inference scheduling with on-site model training. This system uses large models to understand complex long-tail scenarios and perform progressive reasoning, while small models execute specific tasks efficiently. It fundamentally solves the classic trade-off in traditional AI deployment between computing cost and accuracy. The share of projects adopting multimodal and agent capabilities jumped from nearly 0% in 2024 to around 60% in 2025.

X Businesses

In 2025, the Group firmly implemented its“1+X”strategy. Through strategic restructuring of vertical businesses and market-oriented incentive mechanisms, we successfully unlocked asset value and operational vitality of our ecosystem enterprises. We have established a highly efficient collaborative system in which the“mothership (Group)”provides the foundational platform, while the sub-vessels (ecosystem) compete in targeted sectors, realizing a strategic shift from standalone business growth to multiplied ecosystem value.

•Continuous expansion and diversification of the innovation ecosystem: Backed by the Group’s foundational enablement in large models and spatial intelligence, we have built one of China’s most dynamic AI ecosystem matrices, forming a complete closed loop covering upstream core components to downstream vertical applications, including: intelligent driving with cutting-edge world model capabilities and mass-production delivery, high-performance inference GPUs, edge AI chips empowering universal intelligent connectivity, a pioneer in the new category of home AI robots, smart healthcare and smart retail focused on industrial digital transformation, and ACE Robotics representing frontier breakthroughs in embodied intelligence. This diversified portfolio allows SenseTime to fully capture growth potential across all sectors where AI penetrates physical spaces and deeply integrates with the real economy.

•Market-oriented financing validates the growth resilience of“sub-vessels”: Ecosystem enterprises incubated by the Group have achieved smooth progress in private market financing, gaining strong recognition from external investors, including internet industry leaders, top-tier venture capital firms and industrial funds. Leveraging the computing infrastructure and foundational models provided by the Group, these ecosystem entities have drastically shortened the cycle from technology R&D to large-scale commercial deployment.

Business Outlook - For the year ended December 31, 2025

Looking ahead to 2026, the global AI industry will enter its critical stage of development and competition. Rising intelligence density and extreme optimization of unit intelligence costs will further drive an exponential boom in generative AI. SenseTime is fully prepared to prove its strength in this decisive moment with an even more focused business portfolio and a leaner organizational structure.

1. Continued commitment to the native multimodal path to define new heights of intelligence:We will continue to deepen our native multimodal architecture and establish our global leadership in integrated native multimodal large models and spatial intelligence, breaking through limits of artificial intelligence via more efficient pathways, thereby empowering AI systems to their full potential.

2. Capturing the wave of industrial intelligent upgrade in the Agent-native era leveraging our native multimodal capabilities: We firmly believe in the trend of AI evolving from a“productivity tool”to“productivity”itself and will deepen our presence in vertical sectors with a high penetration rate of intelligent agents, including general office, education and marketing. Meanwhile, we will further explore differentiated experiences for a new generation of AI-native applications, achieving explosive growth in both user scale and commercial value.

3. Extreme optimization of unit intelligence costs and full promotion of domestic substitution:Building on the synergistic advantage of the three-in-one“computing power–model–application”strategy, we will continue to reduce inference costs for large models, aiming to make exceptional cost-effectiveness our most solid competitive barrier.

4. Computer Vision: We will consolidate our“profit formula”of CV2.0, and drive large-scale business expansion. At the same time, we will maintain dual engines at home and abroad, and establish the“SenseTime Solution”as a global benchmark for the smart industry.

5. Continuously unlocking ecosystem value from the“1+X”strategy: We will further refine the“mothership + sub-vessels”cluster collaboration model, keep empowering ecosystem entities through synergy, and capture value-added dividends from the intelligent transformation boom across vertical industries.

Source: SenseTime-W (00020) Annual Results Announcement