China AI Startup Pulse

# 2026年：中国AI行业迈入规模化商业落地新纪元，行业格局加速演变 2026年，中国人工智能（AI）行业迎来了前所未有的关键转折点，从技术创新到产业落地，全面迈入规模化商业化的新时代。国产AI在内容生成、多模态智能、Agent系统、Embodied AI（机器人与自动驾驶）以及自主硬件（芯片）等关键领域实现了突破，不仅在国内市场掀起浪潮，也逐渐在国际舞台展现竞争实力。 ## 技术创新与产业落地的全面突破 今年被普遍视为中国AI从实验室走向产业前沿的关键年份。多项技术实现了从科研到产品的转变，推动行业快速规模化落地。 - **多模态内容生成**：字节跳动推出的**Seedance 2.0**，支持视觉、音频及文本的深度融合，极大提升内容创作的效率和品质，成为娱乐、媒体、广告等场景数字化升级的核心引擎。 - **Agent化交互**：字节跳动的**Doubao 2.0**代表“Agent时代”的到来，支持多任务协同和个性化交互，为用户带来更自然、更智能的交互体验。这一模型的发布，标志国产AI在自主、多任务Agent系统方面实现了实质性突破。 - **行业应用深化**：制造、金融、医疗等行业纷纷引入国产AI模型，实现医疗辅助诊断、金融风控、智能制造等商业化应用，逐步构建起完整的产业生态链。 ### 资本市场的热烈表现 行业资本动向同样令人振奋： - **Zhipu AI**在香港完成超过5.58亿美元融资，彰显其在高端模型研发中的雄心。 - **MiniMax**股价翻倍，验证其创新能力与市场潜力。 - **Kimi**获得5亿美元融资，布局模型全产业链。 - **DP Technology**募集1.14亿美元，用于强化工业仿真及算力基础设施。 - **StepFun**在B+轮融资中获得50亿人民币，专注于对话式AI和内容创造。 特别值得一提的是，**Doubao 2.0模型**的推出，正式开启“Agent时代”。数据显示，国产多款自主研发模型已在知名应用平台中占据主导地位，国产模型的国际影响力也在不断扩大，彰显中国在智能交互和多任务Agent系统领域的竞争力。 ## Embodied AI与机器人：商业化进程加快 中国在自动驾驶和机器人产业的突破持续跃升。多个城市的**Robotaxi**已实现“无干预”规模运营，成为行业的重要里程碑。**地平线（Horizon Robotics）**和**小马智行（Pony.ai）**的自动驾驶车辆，在复杂交通环境中的表现日益成熟，安全性和稳定性获得高度认可。 **最新数据显示：** - 多个城市的Robotaxi已实现规模化运营，显著提升出行效率，应对复杂交通场景。 - **“Built for chaos”**设计理念增强自主适应复杂环境的能力，提升交通安全。 - **阿里巴巴**推出的**RynnBrain多模态模型**，支持自主决策、自然交互和自主学习，已广泛应用于制造、物流与服务行业。 - 在国际合作方面，阿里巴巴与OpenAI等全球巨头合作，共同优化机器人交互体验，推动在更复杂环境中的自主适应能力。 在文化娱乐方面，**“功夫机器人”**类短视频在社交媒体走红，展现了高度灵活、精准的动作能力。这不仅体现技术突破，也逐渐渗透到文化娱乐和消费场景中，产业格局日益多元。 自动驾驶与智慧物流的深度融合，以及Embodied AI在城市管理和制造业中的多场景应用，共同塑造了“人机协作”的完整生态体系，极大提升中国在智能交通和机器人产业的国际竞争力。 ## 自主硬件：国产芯片持续突破 硬件自主研发成为行业核心驱动力之一。中芯国际推出的**Zhonghaoxinying 7nm芯片**，代表国内高端芯片制造的重大突破，为大规模训练和部署AI模型提供坚实的算力基础。这一芯片的问世，大幅降低对进口GPU的依赖，增强产业自主能力。 阿里巴巴、华为及中科院等机构也在不断推进自主芯片研发： - **阿里巴巴的“Zhenwu 810E”**芯片，已广泛应用于企业硬件，显著提升模型推理和训练效率。 - 新一代计算架构的开发，使处理速度提升近四倍，为超大规模模型（如**Intern-S1-Pro**）提供有力算力支撑。 数据显示，Nvidia在中国市场的份额由66%骤降至约8%，国产硬件的崛起开始蚕食国际巨头市场份额，逐步实现芯片自主可控。阿里云的**“Scaling Spatial Intelligence”**项目，在空间感知和空间计算方面取得突破，为智慧城市和自动驾驶提供坚实基础。 ## 旗舰模型、生态平台及国际布局持续推进 2026年，国产超大模型不断刷新能力边界： - **GLM-5**由智谱AI开发，采用**7440亿参数的MoE架构**，结合创新的“Slime”训练技术，特别适合复杂推理和长链Agent任务。 - **Kimi的K2.5模型**，声称拥有“万亿参数”，引发行业关注，其高效训练和卓越性能广受期待。 这些模型已在智能制造、金融分析、科研探索等领域实现试点应用，彰显中国在超大规模自主研发模型上的持续突破。 在国际市场方面，国产方案逐渐崭露头角： - Airbnb宣布放弃使用OpenAI技术，转而采用中国本土企业的AI模型以满足其定制化与数据安全需求。 - 关于**Anthropic**指控多家中国AI企业“采矿”Claude模型数据的报道，引发行业关于数据伦理和合规的讨论。 此外，**脑机接口（BCI）**技术在中国快速发展，传感技术与具身交互创新不断推进，为智能人体、增强现实等前沿应用奠定基础。 ## 新动向：多代理平台崛起与国际亮相 - **Astron Agent**：开源多代理AI自动化平台，支持多Agent协作、复杂环境适应和自主学习，成为推动行业自动化的重要力量。相关展示视频显示，其已成为智能自动化的重要工具，并在业界引发热议。 - **国际展会亮相**：在“Automation World Seoul”展会中，中国Humanoid机器人企业频繁亮相，展现出高度智能、良好的交互能力，获得国际关注。这些机器人不仅技术成熟，还在文化输出方面展现出巨大潜力。 - **全球使用统计**：根据**OpenRouter**最新报告，**中国AI模型**首次在全球使用量中超越美国模型，MiniMax等国产模型占据主导地位，显示出中国自主模型在国际市场的竞争力持续增强。 ## 行业面临的挑战与未来前景 尽管行业高速发展，但仍面临多重挑战： - **供应链与合规风险**：在美国出口限制背景下，DeepSeek等企业利用高端芯片推进研发，行业供应链的复杂性和潜在合规问题日益突出。 - **数据伦理与安全**：关于Claude模型“采矿”数据的指控凸显行业在数据安全、伦理和隐私方面的压力。 - **国际压力**：美国不断加强技术出口管制，促使中国企业加快自主创新步伐，推动硬件和软件的自主可控。 中国政府积极推动行业规范，强化安全审查，确保技术的健康、可控发展。未来，行业将重点关注： - **绿色高效模型**：降低能耗和成本，促进可持续发展。 - **自主生态系统**：完善硬件与软件的自主配套，减少对外依赖。 - **国际合作**：推广中国自主AI方案，扩大全球影响力。 - **伦理与安全体系**：构建完善的行业伦理和安全标准，保障技术持续健康发展。 ## 重要新动向：企业估值与市场策略 **Qianxun Intelligence**在短短两年内估值突破**10亿人民币**，成为“百亿俱乐部”新晋成员。尽管市值飞涨，但公司高层表示“不能放松警惕”，持续投入研发以保持竞争优势。这反映出行业对优秀 humanoid 机器人企业的巨大热情和市场期待。 同时，**Alibaba**正大幅降价，推动其AI编码工具普及，旨在“夺占”市场份额，提升平台竞争力。这一策略或将引发行业价格战，促使企业不断优化技术和商业模式。 ## 结语：迈向全球科技新高点 2026年，中国AI行业正处于由创新驱动向规模化落地的关键转折点。自主模型和硬件的持续突破，产业生态的不断完善，使中国从技术探索逐步走向广泛应用，从实验室走向千家万户。在国际环境日趋复杂的背景下，行业领导者积极应对，推动中国AI在全球科技版图中的地位不断提升。 未来，绿色、高效、可控的AI技术将成为行业发展的核心动力。中国的“智慧未来”正逐步成为现实，展现出强大的科技创新能力和产业转型潜力。中国AI，正以坚定的步伐，迈向全球科技的崭新高点。

# China’s AI Ecosystem in 2026: DeepSeek’s V3-0324 and the Drive Toward Self-Reliance Expanded In 2026, China continues its bold march toward establishing itself as a dominant global force in artificial intelligence. The landscape is characterized by cutting-edge innovations, strategic hardware and software developments, and an increasingly resilient open-source ecosystem. Central to this momentum is **DeepSeek’s recent launch of the upgraded V3-0324 model**, a milestone that exemplifies China’s commitment to **AI self-reliance, international competitiveness, and industrial integration**. This development, alongside other breakthroughs across sectors, signals a comprehensive evolution of China’s AI ambitions and capabilities. --- ## DeepSeek’s V3-0324: Elevating Enterprise AI to New Heights Building upon its earlier versions, **DeepSeek’s V3-0324** has garnered widespread attention for its **transformative features** that reinforce China’s position in the global AI arena: - **Revolutionary Long-Term Memory Capability** The model now incorporates a **million-token memory extension**, a **world-first feature** that enables **extended, coherent, and context-rich interactions**. Industry experts like Li Wei highlight that **"the million-token memory sets a new industry standard,"** facilitating AI applications where **long-term contextual understanding is crucial**, such as **scientific research, complex customer interactions, and enterprise decision-making**. - **Enhanced Reasoning and Deep Understanding** The V3-0324 exhibits **notable improvements in nuanced comprehension**, making its responses **more accurate, context-aware, and reliable**. This is particularly vital for **technical support, research collaboration, and industry-specific solutions** where precision matters. - **Faster Inference & Deployment** Thanks to **architectural and software optimizations**, inference speeds have been **significantly increased**. This allows for **real-time virtual assistants**, **decision support systems**, and **interactive enterprise tools**, broadening the AI’s application across **multiple sectors**. - **Cost-Effectiveness & Democratization of AI** The model now **requires fewer computational resources**, lowering entry barriers for **startups, SMEs**, and **large enterprises alike**. This affordability **fosters a vibrant, domestically driven AI ecosystem**, accelerating **adoption and innovation** at all levels. - **Benchmark Performance and International Validation** On renowned benchmarks such as **GLUE and SuperGLUE**, V3-0324 **matches or exceeds** the performance of Western counterparts like **OpenAI’s GPT-4** and **Anthropic’s Claude**. This **credible validation** positions the model as a **viable alternative for enterprise AI solutions**, reinforcing China’s **AI independence**. ### Hardware & Supply Chain Resilience A core pillar of China’s AI strategy is **hardware independence**. Despite ongoing geopolitical tensions, **DeepSeek has trained its latest models using Nvidia’s H200 GPUs** through **domestic partnerships and alternative supply chains**, demonstrating **resilience against export restrictions**. Meanwhile, **Chinese chipmakers** have made significant strides: - **Huawei’s Ascend chips** are now capable of supporting **large-scale model training**, pushing China toward **hardware self-sufficiency**. - **Zhipu AI** recently trained a **major model solely on Huawei chips**, signaling a move toward **integrated hardware-software ecosystems**. - **Domestic chip innovations** from firms like **Moore Threads** are making tangible progress in **reducing reliance on foreign hardware**, though **Nvidia’s H200 GPUs** remain vital for **large-scale training and inference**. --- ## The Broader AI Ecosystem: Innovation, Competition, and Open-Source Leadership China’s AI landscape is buzzing with **flagship models, open-source initiatives, and innovative applications**: - **Z AI’s GLM-5** The recently launched **GLM-5** is a **robust, multi-modal model** tailored for **complex systems engineering** and **enterprise solutions**. Its emphasis on **community-driven development** aims to **foster collaborative innovation** and **market diversification**. - **MiniMax’s M2.5 Series & Kimi K2.5** The **MiniMax M2.5 and Lightning models** continue to **push performance boundaries**, often **approaching Western state-of-the-art accuracy at roughly 1/20th of the cost**. The **Kimi K2.5**, a **trillion-parameter open-source model supported by Alibaba and HongShan**, underscores China’s **dedication to scalable, high-performance open models** as a **counterbalance to Western dominance**. - **Rising Influence of Chinese Open-Source Models** The **OpenRouter platform** reports that **Chinese models**, led by **MiniMax**, **now surpass US models in global usage**, reflecting **international adoption**. The proliferation of **PicoClaw**, a **lightweight, resource-efficient open-source AI agent**, reduces resource requirements by **up to 99%**, enabling **agentic AI on resource-constrained devices**—a **game-changer for democratized AI deployment**. --- ## Multimedia & Agentic AI: Reshaping Content and Society Chinese firms continue to **push the boundaries** in **multimedia AI** and **agent systems**: - **ByteDance’s Seedance 2.0 & Doubao 2.0** ByteDance’s **Seedance 2.0** has demonstrated **highly realistic cinematic video generation**, which has gone viral, garnering **over 19,700 views**, **510 likes**, and **220 comments**. These capabilities **set new standards** for **multimodal content creation**, transforming media production workflows. **Doubao 2.0**, designed specifically for **agentic AI applications**, automates **content production, personalized assistance**, and **multimodal interactions**, significantly **reshaping media and communication industries**. - **Baidu’s Integration of OpenClaw & Multi-Agent Platforms** Baidu has embedded **OpenClaw**, an **advanced AI agent platform**, into its **search ecosystem**, enabling **interactive, autonomous assistance**—such as **scheduling, information retrieval**, and **personalized recommendations**. The platform’s **success** signifies the **mainstreaming of agentic AI**. Additionally, the **Astron Agent** project—a **collaborative, open-source multi-agent platform**—is enabling **complex task automation** across sectors, further **enhancing autonomous capabilities**. - **Humanoid Robots and Robotics** At the **Automation World** exhibition in Seoul, **Chinese humanoid robots** showcased **advanced mobility and interaction skills**, exemplifying China’s **growing leadership in robotics** and their potential for **industrial, service, and research applications**. --- ## Sectoral Deployments & Industry Impact AI’s transformative influence spans multiple industries: - **Payments & Financial Services** **Alipay’s AI payment system** processed **over 120 million transactions in a single week**, demonstrating **agentic automation**—from **smart contracts** to **personalized recommendations**. - **Insurance & Risk Management** **Zhibao Technology** integrated **AI agents** into its platforms, **streamlining underwriting, product design**, and **risk assessment**, resulting in **significant operational efficiencies**. - **Autonomous Vehicles & Urban Robotics** Chinese **robotaxi fleets** are **expanding urban deployments**, with reports indicating **superior urban resilience**—a direct result of **AI systems optimized for unpredictable environments**. --- ## Funding, Challenges, and Strategic Responses Investment in China’s AI sector remains **robust**: - **Recent funding rounds** include **AI² Robotics’ USD 145 million** to support **model development and robotics commercialization**. - Despite **geopolitical pressures and compute shortages**, Chinese firms are **innovating around export restrictions**. **DeepSeek’s models trained via domestic supply chains** exemplify this strategic resilience. However, **challenges persist**: - **Compute resource shortages**—highlighted by **Zhipu AI’s recent share price decline**—underline the urgency for **further domestic hardware breakthroughs**. - **Regulatory and ethical frameworks** are evolving to **ensure privacy, safety**, and **trust**, critical for **long-term sustainable growth**. - **Geopolitical tensions**, especially **US export controls**, continue to pose hurdles, but **China’s strategic alliances and domestic innovations** reinforce its **AI independence trajectory**. --- ## New Developments: Valuations, Robotics, and Market Competition Adding to the momentum, **Qianxun Intelligence** recently achieved a remarkable milestone: its **valuation exceeded 10 billion Yuan within just two years**. Despite this rapid valuation growth, the company **dare not relax**, reflecting **the intense competition and high stakes** in the humanoid robot sector. The **“Hundred-Billion Club” for humanoid robots** has added a sixth member, emphasizing **China’s expanding leadership in robotics**. Simultaneously, **Alibaba** has aggressively **slashed prices for its AI coding tools**, aiming to **capture market share** and **set standards in cost-effectiveness**. This move **intensifies competition** and **accelerates adoption of AI development tools**, especially among **smaller firms and individual developers**. --- ## Current Status & Future Outlook In 2026, China’s AI ecosystem is **remarkably dynamic**, characterized by **innovative models like DeepSeek V3-0324**, **a resilient hardware supply chain**, and **a thriving open-source community**. The **integration of multimodal, agentic, and autonomous systems** is transforming **industries, content creation, and societal interactions**. The **DeepSeek V3-0324’s launch** exemplifies **China’s leadership in enterprise AI**, complemented by **cost-effective models such as Qwen3.5** and **aggressive market strategies by firms like Alibaba**. The **rise of open-source models like PicoClaw** and **multi-agent platforms like OpenClaw and Astron** signals a **paradigm shift towards more autonomous, interactive AI systems**. Despite ongoing challenges—**compute shortages, geopolitical tensions, and regulatory needs**—China’s **strategic resilience, innovation capacity, and strategic alliances** position it to **sustain its trajectory as a global AI leader**. As the ecosystem matures, **China’s AI ambitions are poised to reshape the future of artificial intelligence, both domestically and internationally**.

# China’s GLM-5 and the Long-Context AI Revolution: Breakthroughs, Challenges, and Emerging Security Concerns In the rapidly evolving landscape of artificial intelligence, **Zhipu AI’s GLM-5** continues to exemplify China’s ambitious stride toward creating **long-term, reasoning-capable AI systems**. Building on its technological foundations—**744 billion parameters**, **advanced multi-tool orchestration**, and **groundbreaking memory integration**—GLM-5 is spearheading the **Long-Task Era**, where AI sustains coherent, reasoning-rich interactions over extended periods. Recent developments reveal a complex scene marked not only by remarkable innovations but also by intensifying geopolitical tensions, fierce competition, and emerging security threats that threaten to reshape the future trajectory of China’s AI ambitions. --- ## Technical Breakthroughs and Ecosystem Expansion **GLM-5**’s recent advancements affirm its position as a leading large-language model (LLM): - **Massive Scale and Reasoning Capabilities:** - With **744 billion parameters**, it supports intricate **multi-turn dialogues**, **domain-specific reasoning**, and **multi-tool orchestration**, enabling over **700 tool calls** and **800 context handoffs**. - These features are critical for applications requiring **long-term coherence** and **complex decision-making**. - **Memory and Context Handling:** - Integration of **DeepSeek’s** **million-token memory system** marks a **paradigm shift**—allowing **GLM-5** to **recall and process extensive contexts** with high fidelity. - This enhancement supports **detailed, sustained interactions** across domains such as **document analysis**, **multi-step reasoning**, and **extended conversations**, vital in sectors from legal to enterprise automation. - **Deployment Ecosystem:** - **International Reach:** - Partnered with **Google Cloud’s Vertex AI**, **GLM-5** now serves global markets, enabling multinational corporations to deploy and customize at scale, positioning itself as a strong competitor to Western giants like OpenAI and Anthropic. - **Domestic Integration:** - Within China, it powers **Kuaishou Vanchin** for content moderation and interactive features, **Novita AI API** for enterprise automation, and **Atlas Cloud** for decision workflows. - **Developer Tools and Open-Source Initiatives:** - Tools such as **Kilo CLI** and **Kilo Code** allow developers to fine-tune and adapt **GLM-5** for diverse use cases. - Open-source projects hosted on platforms like **N3** and chat APIs on **N1** foster a vibrant community, democratizing access and spurring grassroots innovation. --- ## Market Impact and Benchmark Successes **GLM-5**’s technological leadership is validated by impressive benchmarks and deployment milestones: - **Benchmark Performance:** - According to **Zai_org**, **GLM-5** ranks **#3 overall on Design Arena** and **#1 among open models**, underscoring its competitive edge against global counterparts. - **Real-World Deployments:** - **Alipay** has integrated **GLM-5** into its AI-driven payment and transaction systems, managing **over 120 million weekly interactions**. Features like **AI Pay** and **Tap!** enable seamless natural language payments and inquiries, illustrating AI’s integration into China’s daily economic activities. - The community’s shared dataset has grown to **1.5 terabytes**, reflecting active collaboration and rapid knowledge exchange among developers and enterprises. --- ## Competitive Landscape: Alternatives and Emerging Models While **GLM-5** maintains dominance domestically, a host of other models are emerging as competitors: - **Alibaba’s Qwen3.5:** Focuses on **long-context workflows** and **AI agent capabilities**, hosted on Alibaba Cloud, offering strong enterprise solutions. - **ByteDance’s Doubao 2.0:** Introduces **expanded context handling** and **advanced agent architectures**, intensifying competition in the Chinese AI market. - **Kimi K2.5:** A highly anticipated **trillion-parameter model**, recently garnered attention following presentations titled **"Kimi K2.5 — The AI Nobody’s Talking About,"** hinting at its potential to rival existing giants. Additionally, **lightweight open-source models** are gaining ground: - **PicoClaw** and **INSANE Chinese $10 Version of OpenClaw** exemplify efforts toward **cost-effective, resource-efficient AI**. - The latter, claiming **99% less memory usage**, underscores a strategic push toward **accessible, democratized AI**, especially vital amid hardware shortages and rising demand. --- ## Geopolitical and Market Challenges Despite technological progress, **Zhipu AI** faces significant hurdles: - **Hardware and Compute Shortages:** - Recently, **Zhipu AI’s stock** plummeted nearly **23% on February 23**, erasing over **HK$70 billion (around US$9 billion)** in market value. - The core concern revolves around **global compute shortages**—a bottleneck threatening to **delay scaling efforts** and **limit international collaborations**. - **Recent reports** reveal that **DeepSeek** trained its **models on Nvidia’s top-tier chips** despite **US export bans**, indicating a strategic effort to **maximize available hardware** and **circumvent restrictions**. - **Geopolitical and Regulatory Pressures:** - The US government’s ongoing **export controls** threaten to **limit Chinese access to advanced AI chips**, complicating **R&D** and **deployment** plans. - **Security concerns** have escalated recently—**Anthropic’s warnings** about **distillation attacks** and **large-scale copying operations** signal heightened risks. - **China’s strict privacy laws**—focused on **data sovereignty**—offer domestic models like **GLM-5** a **competitive advantage** in training and deployment, though they may **hamper cross-border data sharing**. --- ## Recent Security Developments & New Threats Adding to the complexity, recent reports have brought to light **serious security concerns**: - **Anthropic’s Warning:** - In a notable video titled **"You Built A MONSTER!"**, Anthropic warns of **massive Chinese AI copying operations**. They highlight that **Chinese firms have generated millions of queries** with **Claude**—a model developed by Anthropic—using **distillation techniques** that could **distort or extract sensitive data**. - This signals a resurgence of **model theft, data mining, and adversarial attacks**—posing **security threats** and **intellectual property risks**. - **Large-Scale Data and Model Extraction:** - Recent investigations reveal **extensive copying operations**, where Chinese AI companies are **replicating and deploying models** at scale, often **without explicit licensing** or **security safeguards**. - These activities raise **regulatory and national security concerns**, prompting calls for **robust defenses** against **model theft**, **distillation attacks**, and **adversarial manipulation**. --- ## Strategic Implications and Future Outlook **GLM-5’s** trajectory underscores several critical points: - **Heavy Reliance on Hardware:** - The global **compute shortage** remains a fundamental challenge. China's strategy of **maximizing existing hardware**—training on **Nvidia chips** despite export restrictions—demonstrates resilience but also underscores vulnerability. - The **domestic push for alternative hardware solutions** and **open-source models** like **PicoClaw** and **INSANE** aim to **reduce dependency** and **democratize AI access**. - **Data Sovereignty as an Advantage:** - China's **strict data privacy laws** craft a **protected environment** for domestic models, fostering **trusted data pools** and **secure training environments**—a potential strategic edge amid international restrictions. - **Open-Source and Grassroots Innovation:** - The rise of **cost-effective open-source models** positions **bottom-up innovation** as a key driver, enabling **smaller firms and research institutions** to participate actively in AI development. - **Security and Regulatory Measures:** - The recent **distillation attacks and large-scale copying reports** emphasize the need for **robust security protocols** and **regulatory oversight** to prevent **model misuse** and **data breaches**. **In conclusion**, **GLM-5** exemplifies China’s **ambitious AI vision**—a convergence of **technological mastery**, **ecosystem resilience**, and **strategic navigation** through geopolitical and market uncertainties. As the **Long-Task Era** matures, **GLM-5’s evolution** and its successors will likely define the future landscape of AI—balancing **powerful reasoning capabilities** with **security and ethical considerations**. The coming years will be pivotal in determining whether **Chinese AI** can sustain its leadership amid mounting challenges, and how **long-context, reasoning-capable AI** will reshape industries globally.