Inside Z.ai's secret model drop, Seedance 2.0's viral moment, and China's billion-yuan AI spending war.
China’s AI sector just had one of its most intense weeks of the year. A year ago, DeepSeek proved that timing could turn a model release into a global event. This Spring Festival, China’s AI labs responded in force.
In this week’s newsletter, we are gonna cover quite a lot of news:
-Z.ai quietly climbed to the top of OpenRouter’s leaderboard under a codename (Pony Alpha).
-ByteDance-backed SeeDance 2.0 pushed video AI closer to director-level control.
-Kuaishou’s Kling embedded generative tools deeper into platform distribution.
Meanwhile, capital markets reacted — and competition escalated into a new round of spending pressure. In this issue, we break down what actually changed — and why 2026 may hinge less on benchmark scores and more on economic resilience.
Welcome to Voice of Context Weekly. This is a new look of our weekly briefing covering the stories, signals, and strategic shifts shaping China’s tech industry, with a focus on AI, capital markets, and the U.S.-China competitive landscape.
Most of what you’ll read here is available nowhere else in English. We go beyond translated headlines. Look for our VoC Channel Checks throughout each issue: firsthand interviews and insights from industry insiders, operators, and investors.
INTERVIEW: Z.ai’s Quiet Test: From Anonymous Drop to Subscription Shift
Z.ai (智谱, 2513.HK) pulled off one of the more creative model launches in recent memory. The company secretly uploaded its new flagship, GLM-5, to OpenRouter under the codename “Pony Alpha,” told no one, and waited. Within days, the anonymous model topped OpenRouter’s popularity charts, processing 60 billion tokens per day. Speculation swirled: Was this DeepSeek V4? Claude Sonnet 5?
It was neither. On February 11, Z.ai confirmed the model was GLM-5. The stock surged 28.68% on the next trading day, pushing the company’s market cap to roughly RMB 150 billion (~$20.5 billion), the highest valuation of any listed LLM company in China.
VoC Channel Check: We sat down with Zixuan Li, head of z.ai, hours after the GLM-5 reveal. The full English video interview is here:
Here are the key takeaways:
Instead of seeking domestic validation first and exporting credibility later, Z.ai used global developer verification as a signal amplifier — letting overseas performance data circulate back into China’s capital markets.
In today’s AI cycle, foreign developers’ attention is no longer just reputation. It is liquidity.
Beyond Model Strength: Scenario Awareness
Z.ai’s leadership does not frame GLM-5 primarily as an infrastructure breakthrough. In our interview, what stood out was their emphasis on real-world use cases.
The company highlighted its deep collaboration with large platforms such as ByteDance, using live scenario data to iterate quickly. Internally, they operate what they call a “GM Coding Plan,” collecting both positive and negative case feedback at scale to refine performance.
This signals a different competitive thesis.
While some Chinese labs pursue scale and compute intensity, Z.ai is positioning around scenario awareness — the ability to adapt to real user workflows faster than competitors.
In a generation cycle that now compresses to roughly three months, responsiveness may matter as much as parameter count.
Coding as Defensive Consolidation
GLM-5’s strongest traction is in coding.
That is not accidental.
Coding remains the clearest path to monetizable PMF among foundation model companies. Enterprise pilots are slow. Consumer chat is volatile. Coding, by contrast, embeds directly into developer workflow.
But Z.ai’s move here is not simply product iteration. It is structural.
The company introduced a subscription-based Coding Plan — shifting API usage away from pure token metering and toward a flat-fee model. Tokens, in their framing, should function like electricity: an invisible infrastructure cost rather than a user-facing constraint.
This represents one of the first serious attempts by a Chinese model company to “ToC-ify” API access, and they are seeing early success.
This does two things:
1. It locks in workflow dependency rather than episodic experimentation.
2. It decouples revenue from raw token volatility in an increasingly competitive price war.
In a market where token prices collapse quickly, subscription models are a form of revenue stabilization.
The Hidden Constraint: Compute Saturation
Perhaps the most revealing part of the conversation was not about features, but about capacity.
Z.ai acknowledged a structural dilemma facing many model labs:
When one generation succeeds, it consumes compute at scale.
But rapid iteration requires freeing capacity for the next generation.
Meanwhile, new competitors can launch improved models unexpectedly, forcing repricing and triggering user migration.
In other words, success creates its own bottleneck.
Compute allocation becomes not just a technical decision, but a financial one. Raise prices too early, and you lose users. Delay upgrades, and you lose momentum. Upgrade too fast, and you cannibalize your own revenue base.
This dynamic — supply shock inside the model cycle — is rarely discussed publicly, yet it may define the economics of the sector.
A Capital Market Experiment in Real Time
Z.ai occupies a unique position among Chinese AI labs. With public market exposure, it operates under valuation pressure uncommon among earlier-stage peers.
GLM-5’s anonymous debut, coding push, and subscription shift together suggest a coordinated attempt to prove two things simultaneously:
That it can compete technically at global benchmark levels.
That it can convert model strength into recurring revenue before the next frontier leap reshapes the landscape.
If DeepSeek’s breakout was about technical surprise, Z.ai’s moment may be about economic stabilization.
And in 2026, that may be the harder challenge.
ByteDance’s Seedance 2.0 Just Broke AI Video. Now Comes the Hard Part.
ByteDance dropped Seedance 2.0 last weekend, and the internet lost its mind. Generated clips flooded Douyin, WeChat, and X simultaneously. Foreign users scrambled for Chinese phone numbers to register on ByteDance’s Jimeng AI (即梦) platform.
The hype is grounded in a genuinely new capability. Seedance 2.0 accepts up to 12 input files across four modalities: images, video clips, audio, and text. An @ mention system lets users assign each file a specific role: this image is the character, this clip is the camera motion reference, this audio track sets the rhythm. The model then composes a multi-shot sequence with synchronized sound, lip-sync in eight-plus languages, and physically plausible motion. One tester produced a 60-second anime short in 15 minutes, no re-rolls needed.
This video was generated from the simple prompt: “Bruce Lee, dressed in his iconic yellow outfit, faces Mike Tyson in boxing gear inside an MMA octagon, trading blows back and forth with neither side gaining the upper hand.”
VoC Channel Check: A-share media and AI stocks surged in response, with multiple names hitting their daily limit-up. Kaiyuan Securities analyst Fang Guangzhao called Seedance 2.0’s test results “stunning” and wrote that it may mark a “singularity moment” for AI in film and television. Orient Securities argued that video generation has entered an era of “precise, dashboard-level control,” with applications in AI comics, animation, and short drama production poised to scale first.
However, this launch also exposed a raw nerve. Tech creator Tim (潘天鸿), who runs the popular “Film Storm” (影视飓风) channel, discovered that Seedance 2.0 could generate videos with his voice and likeness unprompted, simply because his public videos were in the training data. ByteDance quickly suspended the real-person reference feature.
Meanwhile, Kuaishou also shipped Kling 3.0 (可灵) in the same window, targeting an entirely different market. Where Seedance optimizes for viral short-form content at low cost, Kling 3.0 goes after professional film production: higher resolution, more realistic skin and muscle rendering, premium pricing. Dongwu Securities analysts summed up the split: “Seedance is for storytelling at scale. Kling is for cinema-grade craft.”
Kimi’s Overseas Revenue Now Tops Its Domestic Business. Here’s How.
Moonshot AI disclosed that Kimi’s overseas revenue has surpassed its domestic revenue, with global paid users growing fourfold. For a Chinese AI startup that most people outside the developer community have never heard of, those are striking numbers.
The breakout traces back to a single product decision. Kimi K2.5, released January 27, is not a benchmark champion. But it turns out to be unusually reliable at the thing that matters most for agent workloads: staying on task. Developers report that K2.5 maintains constraint fidelity over tens of thousands of tokens, doesn’t hallucinate mid-chain, and doesn’t improvise beyond its instructions. In chatbot terms, that’s boring. In agent terms, it’s gold.
The tipping point came when OpenClaw, the open-source agent framework that blew up to 20,000+ active deployments in its first weeks, listed Kimi K2.5 as a recommended model. Developer forums and GitHub filled with deployment guides. Kimi says its token consumption on the platform surpassed Google’s Gemini 3 Flash.
Still, there’s a ceiling in plain sight. On February 5, Kimi posted on Weibo that it is “genuinely short on GPUs” and publicly asked for sourcing help. In a market where Alibaba, ByteDance, and Tencent deploy proprietary clusters at will, compute constraints are an existential bottleneck for startups. Kimi’s long-context lead in 2024 evaporated within months once big tech caught up. Whether its agent-first positioning holds up longer is the central question for the company’s next chapter.
China’s AI Giants Just Spent Billions on Spring Festival Red Envelopes. Will Any of It Stick?
China’s biggest tech companies turned the Spring Festival into an AI user-acquisition arms race. Alibaba committed RMB 3 billion (~$410 million) to a “Spring Festival Treats” campaign for its Qwen chatbot, offering cash red envelopes and free meals, rides, and shopping credits across Taobao, Ele.me, and Gaode Maps. Baidu put up RMB 500 million in red envelopes tied to its Wenxin assistant. ByteDance’s Doubao landed the exclusive AI partnership with CCTV’s Spring Festival Gala, the most-watched broadcast on Earth.
Then Tencent shot itself in the foot.
Tencent launched a red envelope giveaway for Yuanbao, its AI chatbot, inside WeChat, the super-app used by over a billion people in China for messaging, payments, and daily services. The mechanic was simple and addictive: share a digital red envelope with a friend, and every time they claimed it, you earned another lottery chance, up to 30 times. In Chinese culture, red envelopes (hongbao) are a traditional way to gift money during holidays, and their digital version on WeChat is a core feature of the platform.The campaign went viral almost instantly. But WeChat’s own moderation team, which operates independently from Tencent’s AI division, flagged it for violating the platform’s anti-spam sharing rules and restricted it to text-only “passcode” sharing.
Tencent’s stock dropped 3.96% that day. The irony was hard to miss: the same company that owns both WeChat and Yuanbao had its biggest AI marketing push shut down by its own platform. CEO Pony Ma had pitched the campaign internally as a chance to recreate WeChat’s legendary 2015 red envelope moment, the viral hit that turned WeChat Pay from a feature
VoC Channel Check: CICC (China International Capital Corporation, one of China’s top investment banks) internet research team framed the broader dynamic bluntly: the chatbot price war is a land grab, not a business model. None of these apps have proven they can retain subsidy-acquired users. The real question, CICC argues, is whether the chatbot phase is even the right game to win. The industry’s center of gravity is already shifting toward agents, where product stickiness comes from utility rather than incentives, and where startups like Kimi and frameworks like OpenClaw are setting the pace ahead of the giants.
For now, billions of yuan have been spent. The downloads are up. Whether any of it converts to lasting engagement is the question that won’t be answered until the red envelopes stop.