China's Tech Giants Push Into Robotics, Video and AI Agents During Lunar New Year

A Holiday Launch Wave Reaches Beyond Chatbots

Alibaba's $434 million promotional campaign, Tencent's 1 billion yuan in holiday giveaways and ByteDance's exclusive AI cloud partnership with China's state broadcaster have turned the Lunar New Year period into a major contest for artificial intelligence. Alongside those marketing campaigns, Alibaba, ByteDance and Kuaishou have released advanced systems for robotics, cinematic video and software that can carry out tasks with limited human direction.

As of February 20, 2026, the coordinated wave of releases shows how China's technology companies are combining new models with the services people already use to watch videos, shop and order food. While American markets focused on announcements from Anthropic and OpenAI, Chinese companies presented systems intended for practical use across several different industries.

The timing gives the launches a large potential audience. Lunar New Year is China's most significant consumer occasion, and technology companies traditionally use the CCTV Spring Festival Gala to showcase innovations to hundreds of millions of viewers. ByteDance's partnership with the broadcaster places its AI cloud services within that holiday competition.

The products also show the breadth of what industry insiders call the physical AI frontier. The term describes systems that reach beyond answering questions in a chat window: robots handling objects, tools generating realistic media and software executing complicated workflows. These applications require different abilities, but all move AI closer to completing practical work.

Google DeepMind chief Demis Hassabis has assessed that Chinese models trail Western competitors by only months. The latest releases suggest that competition is increasingly about how cheaply models run, how reliably they complete tasks and how easily companies can put them into existing products.

Alibaba Gives Robots a Memory of Their Tasks

Alibaba's DAMO Academy introduced RynnBrain, an open source robotics model designed to give machines a better understanding of space and time. Conventional vision systems can process individual images without retaining enough information about what happened before. RynnBrain incorporates temporal and spatial awareness, allowing a robot to remember when and where events occurred.

A robot using that memory can track progress through a task involving several steps and resume an activity after an interruption. For a machine retrieving something from a refrigerator, recognizing the object is only part of the job. It must also locate it, plan how to reach it and keep track of what it has already done.

Demonstration videos showed robots with simple grippers counting oranges, selecting specific fruits from piles and retrieving milk from refrigerators. These familiar activities require object recognition, a map of the surrounding space, planning for movement and memory of the sequence of actions.

Adina Yakefu, a researcher at Hugging Face, explained how the model's memory changes the way a robot responds to its surroundings:

Instead of simply reacting to immediate inputs, the robot can remember when and where events occurred, track task progress, and continue across multiple steps. This makes it more reliable and coherent in complex real-world environments.

RynnBrain uses a mixture-of-experts architecture and activates only 3 billion parameters during inference. Parameters are the numerical values a model uses to produce its responses; inference is the stage when a trained model performs a task. A mixture-of-experts design calls on selected parts of the model, helping keep the computational cost manageable.

Alibaba says RynnBrain outperforms Google's Gemini Robotics-ER 1.5 and Nvidia's Cosmos-Reason2 across multiple benchmarks. Its release on GitHub and Hugging Face puts it alongside the company's wider open source strategy, undercutting competitors that charge for comparable capabilities delivered through cloud services. Alibaba is seeking a foundational intelligence layer for industrial automation, logistics and healthcare applications.

Seedance Moves Video Generation Toward Production

ByteDance's Seedance 2.0 and Kuaishou's Kling 3.0 extend the holiday launch wave into video production, directly challenging OpenAI's Sora. Their improvements move video generated from text closer to a professional production tool, with analysts describing the emerging capability as a platform suited to industrial standards for creative content.

Seedance 2.0 processes text, images, audio and video inputs together. It can generate cinematic sequences lasting up to 15 seconds, with consistent characters, realistic physics and synchronized audio in multiple languages. Combining those inputs gives creators several ways to specify the result instead of relying on a written description alone.

ByteDance integrates the system into its Dreamina suite within CapCut. That connection positions Seedance for advertisers, social media creators and film producers who need video tools within an established production workflow.

Billy Boman, a creative director based in Stockholm who produces content generated with AI, described the change from the short, limited clips available in 2023 to today's more capable systems:

Back in 2023 it was difficult to get someone to run or to walk. Any type of realism was limited to very short clips, everything was very slow, bad textures, no skin textures, lacking detail. Now the script has flipped. Now I can do anything. It has been nothing short of exceptional, the technological advancements.

Character consistency is particularly useful for storytelling. A person appearing in one shot needs to remain recognizable in the next, while movement, textures and sound must fit the scene. Improvements across those areas explain why the new systems are attracting attention beyond experimental clips.

A Voice Feature Brings Consent Into Focus

ByteDance suspended Seedance's controversial Face-to-Voice feature after it exposed concerns about consent. The feature could reconstruct a person's specific voice and speaking style from a single photograph, without audio samples or permission.

Chinese technology reviewer Tim Pan demonstrated the capability on February 10. After the system cloned his voice from an image alone, he described the experience as terror-inducing. The demonstration brought attention to how quickly a creative tool could reproduce an identifiable person's likeness and voice.

A viral video depicting a fabricated fight between Tom Cruise and Brad Pitt also drew formal complaints from the Motion Picture Association over massive infringement. Together, the voice demonstration and the fabricated celebrity scene placed safeguards and intellectual property rights at the centre of the product's rollout.

ByteDance's suspension addresses one feature, while the wider challenge concerns the relationship between increasingly realistic output and permission to reproduce people or protected material. Creators gain more control over a scene, but the same capabilities make consent and ownership more consequential.

Kling Targets Consistent Characters and Precise Shots

Kuaishou's Kling 3.0 focuses on professional storytelling, including keeping subjects consistent across sequences containing multiple shots. It also stresses physical realism in complex scenes and native audio generation in multiple languages and dialects.

The model extends the duration of a single generation to 15 seconds, matching the maximum clip length described for Seedance 2.0. Kling allows creators to specify the duration, framing and camera movement of individual shots. Those controls give a director more influence over how a sequence unfolds.

Kling also supports batch group image output for storyboarding. A storyboard lays out scenes before production, helping creators plan the progression of a story. Producing related images together connects the planning stage with the eventual video.

As of February 20, Kling 3.0 is available to paying subscribers, with a public release described as imminent. No exact date was given for that wider release. Kuaishou's share price has risen more than 50 percent over the past year, reflecting investor confidence that content tools using AI will deepen engagement and create revenue in short video services.

Kuaishou's pitch is therefore about continuity as much as image quality. Film demonstrations, cinematic narratives and other carefully produced content need characters, environments and actions to remain coherent from one shot to another.

Open Models Extend the Competition to Coding and Agents

Zhipu AI released GLM-5, claiming that it approaches Anthropic's Claude Opus 4.5 in coding benchmarks and surpasses Google's Gemini 3 Pro in specific tests. MiniMax introduced its updated M2.5 open source model, with improved tools for autonomous agents.

An AI agent is a system that can execute a workflow involving several steps with minimal human intervention. Scheduling, research and software deployment are examples. The ambition is to move from a model that tells someone what to do to a system that can carry out the sequence itself.

Chinese laboratories are accelerating the open source distribution of large language models and agent systems. Their approach contrasts with the closed commercial models used by some Western AI leaders. Publishing model weights, training code and datasets helps developers adopt the technology, provide feedback and build new applications.

Model weights are the learned numerical values that shape a model's responses. Making them available gives developers a way to work with the model directly. Open distribution also supports efforts to work around the computational constraints imposed by U.S. export controls on advanced semiconductors.

Alibaba's Qwen family overtook Meta's Llama as the most downloaded model series on Hugging Face in 2025. Teams originating in China currently occupy eight of the platform's top ten trending positions. Those measures show how widely Chinese models are reaching the developer community.

Qwen and Doubao Bring Agents to Large Audiences

Alibaba has expanded its portfolio with Qwen 3.5, designed for the agentic AI era. Its visual agent capabilities allow independent action across mobile and desktop applications, connecting the model's ability to interpret a screen with the steps needed to complete a task.

Alibaba claims Qwen 3.5 is 60 percent cheaper to operate and eight times better at processing large workloads than its predecessor. Cost matters when a model must repeatedly inspect information, choose an action and continue through a workflow. Lower operating costs can make those repeated uses more practical.

ByteDance concurrently released Doubao 2.0, upgrading China's most widely used chatbot application, which is approaching 200 million users. That audience gives the company a large existing base for introducing new AI capabilities.

Alibaba and ByteDance are combining technical development with distribution. A capable model becomes more useful commercially when it is placed inside applications that people already open, rather than requiring them to adopt a separate service.

Three Companies Pursue Different Commercial Goals

Huachuang Securities identifies distinct strategies among ByteDance, Kuaishou and Alibaba. ByteDance is building efficiency infrastructure: accessible tools with low costs and broad capabilities, comparable to a more advanced CapCut. The aim is to reduce content production costs across the web and support its ecosystem through large volumes of output.

Kuaishou concentrates on professional storytelling. Its priorities include physical simulation, realism in complex scenes and character consistency for film demonstrations, cinematic narratives and high quality content that requires continuity.

Alibaba is pursuing integration with e-commerce. Its Qwen-Image-2.0 strengthens product digitization and commercial workflows, linking image capabilities to the practical needs of online selling.

Alibaba's holiday promotion generated more than 120 million orders in six days through coupon distribution powered by AI and integration with food delivery. Alongside the $434 million campaign, that result shows how the company is connecting AI interaction with purchases and orders.

The companies share a launch window, but their commercial destinations differ. ByteDance seeks cheaper, faster production across its content ecosystem. Kuaishou seeks more controllable storytelling. Alibaba connects its models to shopping and other transactions.

More Control Could Reshape Content Production

Huachuang Securities' analysis describes a broader change in the supply of creative content. As video systems become more controllable, creators can spend less time repeatedly generating clips in search of a usable result.

That repeated trial process is sometimes called a gacha system: a creator keeps generating content until a satisfactory version appears. More predictable production moves the cost of creating another video closer to the computational expense of running the model.

Analysts anticipate substantial effects on marketing agencies, e-commerce service providers, manhua, or Chinese comics, and the short drama industry. These businesses all depend on producing visual material, often in large quantities or within tight schedules.

Lower production barriers may also concentrate intellectual property value around leading content. When making images and videos becomes easier, recognizable characters, stories and other valuable creative properties can account for a larger part of what distinguishes one production from another.

Seedance's consent controversy and the Motion Picture Association's complaint sit directly beside that commercial prospect. The same improvements that make production more useful also make control over voices, likenesses and protected material a pressing business issue.

A Short Timeline of the Competitive Shift

The Lunar New Year launches follow changes in model adoption and access to foreign AI services. Several dates help explain how the current competition developed:

• 2023: Billy Boman recalls difficulty generating realistic walking or running, with useful results limited to very short clips and weak textures.
• July 2024: OpenAI banned API access for Chinese developers.
• 2025: Alibaba's Qwen overtook Meta's Llama as the most downloaded model series on Hugging Face.
• Late 2025: DeepSeek's viral rise demonstrated that Chinese laboratories could match Western model performance at a fraction of the computational cost, accelerating competition across the sector.
• February 10, 2026: Tim Pan demonstrated Seedance's ability to reproduce his voice from a photograph.
• Lunar New Year period, February 2026: Alibaba, ByteDance and Kuaishou launched models spanning robotics, video and agent systems, alongside major holiday promotions.

Microsoft also terminated Azure OpenAI services for individual developers in mainland China, while account suspensions have been reported on Anthropic's Claude. These restrictions have intensified domestic development efforts as Chinese companies seek models and services they can operate within their own ecosystems.

Efficiency and Access Define the Global Contest

Demis Hassabis's assessment that Chinese models trail Western competitors by only months provides context for the holiday launches. Chinese laboratories are optimizing models for locally available hardware and developing domestic chip capabilities, with the gap narrowing fastest in inference efficiency and applications that combine different types of input.

China accounts for more than 70 percent of global AI patent filings. U.S. companies continue to lead in foundational architectures and citation impact, measures connected to the underlying designs of models and the influence of research.

U.S. export controls on high bandwidth memory chips and advanced manufacturing equipment have complicated China's progress without stopping it. Huawei is among the companies advancing semiconductor production lines intended to support greater domestic self sufficiency.

The competition now extends beyond benchmark scores. RynnBrain's selective use of its architecture, Qwen 3.5's claimed operating savings and the integration of video models into creative services all point toward the importance of inference efficiency and access to established audiences.

China's robotics, video and agent launches also show why leadership can vary by domain rather than belonging entirely to one country. A company may compete through robot reasoning, another through cinematic control and another through the ability to turn a conversation into a commercial transaction. The Lunar New Year offensive brings those different approaches into the same highly visible moment.

The Essentials

• Alibaba, ByteDance and Kuaishou used the Lunar New Year period to release systems for robotics, video production and autonomous workflows.
• Alibaba's RynnBrain incorporates spatial and temporal memory, activates 3 billion parameters during inference and is available on GitHub and Hugging Face.
• Seedance 2.0 and Kling 3.0 generate clips lasting up to 15 seconds, with improvements in consistency, realism and multilingual audio.
• ByteDance suspended Face-to-Voice after concerns about reproducing a person's voice from a photograph without consent.
• Zhipu AI's GLM-5, MiniMax's M2.5 and Alibaba's Qwen releases extend the push into coding, open models and AI agents.
• Alibaba's promotional campaign generated more than 120 million orders in six days; ByteDance's Doubao chatbot is approaching 200 million users.
• Google DeepMind chief Demis Hassabis assesses the gap between Chinese and Western models at only months, while efficiency and product integration are becoming central competitive factors.