Hi all, there are six days until Lunar New Year eve, and most of East Asia is on holiday, except for the AI labs. Like one by one is rushing to get something out.
Two releases dominated China’s AI chatter this past week. Let’s take a look.
The first was ByteDance’s Seedance 2.0, a new multimodal video model that can take text, images, audio, and video as inputs and generate clips that would normally require real production and editing skills. Early testers have praised its lifelike characters and finer control over specific parts of a clip, representing a significant improvement over previous text-to-video models.
Twitter has been flooded with examples, see [here], [here], and [here]. Period dramas, action blockbusters, Crystal Liu (刘亦菲) as your doting girlfriend, and feeding you ice cream. It seems capable of doing it all.
That said, as AI slop-like glitches appear to have resolved, it is also becoming clear that last-mile touch-up tools will remain necessary. You can go from concept to something surprisingly coherent, but getting from “good” to “exactly what I want” still requires post-production and human taste.
The reason this model’s release was so heavily hyped is that ByteDance is uniquely positioned here due to its distribution and data. The company has claimed more than 4 billion monthly active users across its suite of apps, so anything it embeds in its creation workflows can ripple across the world’s largest short-video ecosystem, truly enabling individual creators and potentially creating a new AI-native film vertical.
The second was Pony Alpha, a mystery model that appeared on OpenRouter with almost no branding and immediately set off a global guessing game.
The name “Pony” fueled speculation. Lunar New Year 2026 begins on February 17, 2026, and it is the Year of the Horse, which makes “Pony” read like a deliberate wink if the builder is Chinese.
Many of the online speculations first jumped to Tencent being the lab behind the stealth model because the founder’s name is Pony Ma, but most serious speculation gravitates toward labs like DeepSeek or Z.ai as they’ve both hinted at dropping something around the lunar new year, unless Tencent’s Hunyuan team has quietly taken a major leap with Yao Shunyu’s arrival.
Let’s dive deeper…
ByteDance turns Seedance into a content flywheel
Let’s start with what ByteDance has already been explicit about. On its own Seed team site, ByteDance describes Seedance 1.0 as a model that supports multi-shot video generation from text and image, capable of generating 1080p videos with smooth motion and cinematic aesthetics. (ByteDance Seed) That is a very particular positioning. Multi-shot is not a spec for fun. Multi-shot indicates that the model is trained on narrative structure rather than on individual clips.
Last year, we wrote a deep dive on ByteDance and Seedance’s role.
TechNode reports that Seedance 2.0 uses a dual-branch diffusion transformer to generate visuals and native audio simultaneously, and that users can input text or upload images and get multi-shot sequences at 2K resolution in under 60 seconds. It also states that the model can produce coherent multi-scene content from a single prompt and that ByteDance has already suspended a feature following scrutiny of privacy and misuse risks.
The business implication is simple. This is not primarily competing with Hollywood. It is enabling a new vertical, AI-native content, where the unit is not a two-hour film but minutes produced per day, tested across variants, localized quickly, and distributed through feeds. And what it can do is that its peers, Kuaishou’s Kling or OpenAI’s Sora, Google’s Veo, cannot do, is what it enables economically.
My understanding of the steps of the attention economy is largely split into 1/ conceptualization, 2/ creation, 3/ production, 4/ (post-production), 5/ distribution, and thus 6/engagement 7/ retention, and finally 8/ monetization.
What ByteDance has that no one else can compete with is: 1) the likely largest pool of video data to train on; 2) one of the largest content-platform distributions (Meta-level); 3) a very mature monetization ecosystem.
Now, with the Seedance production tool, these create a double- or triple-flywheel: more input yields better output, which in turn drives more engagement and, ultimately, more business activity, and the virtuous cycle continues.
Since distribution is a scarce asset. When a company with that kind of reach ships an objectively SOTA creation tool, it does not need to “go to market.” Adoption need not be a battle.
This all falls within the backdrop of the rise of AI micro-dramas. Seedance lands inside an arms race that has already been underway. Kuaishou, the other Chinese short-video giant, reported average MAUs of about 715 million and average DAUs of about 409 million on its core app in Q2 2025, and it explicitly called out growth in its Kling AI business in the same financial release.
Most recently, I met with their representative in November and wrote about their strategy to double down on partnerships with film studios and individual film creators. But this isn’t just a lab-to-lab race in technology; it's two distribution platforms competing to own the creator stack and monetize downstream inventory.
Now, I think there are two overlooked points that make Seedance more important than the flashy demos and what mainstream discussions land.
First, these models need input. Not just prompts, but scaffolding: scripts, storyboards, reference images, reference clips, edits, character sheets, voice direction, and continuity constraints. The near-term future is not “humans disappear.” Because ByteDance has one of the largest pools of video creators in any ecosystem, what happens instead is “humans move upstream”: from filming every second to feeding the machine the right reference material and intent so the output becomes reliable. That shift creates demand for a different kind of labor and a different kind of IP pipeline. And who owns more IP than ByteDance? (Disney? Universal? but again, it’s not going for those.
Second, ByteDance’s deepest edge is video data, especially video data tied to engagement feedback loops. This is the structural problem with the “new lab will win multimodality” narrative. A player like MiniMax can be excellent at model training, but it does not naturally possess the same always-on, engagement-labeled video corpus that a short-video platform accumulates by default. ByteDance does, and it is refreshed daily.
The business model implication is straightforward. Cheaper minutes increase supply. Higher supply increases inventory. More inventory expands monetizable attention, whether through ads, commerce, creator monetization, or some combination. The winner is not the creator with the best prompt. In this vertical, the winner is the platform that can turn faster creation into more learning per day and more monetizable minutes per day.
That said, it also entails obvious externalities and risks. If a model can infer personal voice characteristics from limited inputs or generate high-fidelity content with consistent characteristics, the surface area of deepfakes and fraud expands. Today, Seedance warned users against uploading their own faces, and TechNode reported that ByteDance quickly suspended a sensitive feature, a reminder that deployment is now colliding with policy.
Pony Alpha: launch strategy becoming more and more wild
If Seedance is about industrializing output, Pony Alpha is about industrializing adoption. In other words, Seedance is a distribution story, and Pony Alpha is a credibility story.
Rewind. The Manus launch story took on a life of its own. It shipped the product, sent out limited trial invites, the launch video went viral, and then the invite codes were sold on black markets for thousands of dollars. Intentional or not, AI product launches have, let’s say, been creative.
On Feb. 6, a mysterious model was dropped. A model that appeared out of thin air with no brand recognition, and usage surges when developers find it unusually capable in real workflows.
OpenRouter lists Pony Alpha as a “cutting-edge foundation model” created on Feb 6, 2026, with a 200,000-token context window, priced at $0 per million tokens, and positioned as strong in coding, agentic workflows, reasoning, and roleplay. OpenRouter also shows the model doing roughly 13.9 billion prompt tokens and 548 million completion tokens per day at the height of the surge.
Kilo, which promoted Pony Alpha to coders, framed Pony Alpha as a specialized evolution of one of the most beloved open-source models from a global lab, while keeping its lineage undisclosed.
This all seemed very much anti-model launch in a world where everyone’s benchmark-maxxing and trying to outcompete on noise when models drop, this feels almost indie. Give away inference, harvest real-world tasks, learn where the model fails, and create a social frenzy that does your marketing for you.
There are many critiques these days that say benchmarks are easier to game, leaderboards are easier to overfit, and “SOTA” has become a meaningless term. A stealth drop flips the trust model. You let users test first, and the internet supplies the verdict.
Anyway, like many others online, my initial hunch was that it was DeepSeek, but the more I thought about it, the less I think it is. The whole anti-launch launch feels uncharacteristic for them. They are the classic academics who publish papers with models.
Likely Chinese, but not Tencent
About the timing. The Lunar New Year officially begins on February 17, 2026, and 2026 is the Year of the Horse. The symbolism feels like a wink if the codename came from a Chinese team. Another thing is, it’s hitting all the benchmarks, and who’s better at 应试教育 examinations? To un-PC overgeneralize, the Chinese, thanks to the heated debate regarding Chinese education on X recently.
Then there is the timing: we know DeepSeek has said they will release something around CNY this year, just as they did two years ago with V3. There are also rumors that Pony Alpha is Zhipu’s next-generation GLM-5 model, as jokes have said that if you ask who it is, Pony replies Zhipu (Z.ai), but then replies Anthropic. An inside joke about GLM’s ‘Claude-like’ lineage. On top of all this, there is speculation that, not long ago, Zhipu’s founder and chief scientist, Tang Jie, also said, “the new GLM is coming soon.”
The above rumors and the fact that Tang Jie just attended a gathering at one of Beijing’s innovation centers where the Chinese President met with industry leaders, are fueling the market reaction in Zhipu’s stock, which has seen a noticeable uptick since Pony Alpha’s drop, nearly tripling its IPO price.
So where does Tencent fit in? Mostly as a meme. The founder’s nickname is Pony, yes, but that is not a model lineage. Tencent’s Hunyuan has not been the default pick for frontier coding and reasoning in the way Pony Alpha’s early testers describe.
The only credible Tencent scenario is organizational change, and maybe if someone like Yao Shunyu truly re-architected the training stack and sped up iteration, the market could be behind the curve. But absent that, Z.ai or DeepSeek remain more likely choices because they align better with the observed behavior and with how China’s top labs have been choosing distribution.
I guess we’ll need to wait and see as the Chinese AI labs battle it out to celebrate the arrival of the Year of the Fire Horse. But what makes me think it’s likely a Chinese model lab is that there is more material about this in Chinese WeChat blogs than on X. Making me think it is definitely Chinese. And what we know for sure is that Pony Alpha demonstrates a new distribution playbook for frontier models and proves again that the product matters more than the brand.
Implications for the industry
First, it proves again that the model is increasingly a feature, not the product. Seedance’s real advantage is not a single release. It is ByteDance wiring generation into a loop and its ecosystem that includes creator tooling, distribution, feedback, and monetization. WE’VE BEEN WRITING ABOUT ECOSYSTEMS SINCE THE BEGINNING OF AI PROEM.
Second, AI is not only a supply shock. It is also a demand shock for inputs. Seedance requires reference material and human-created scaffolding to generate reliably, and Pony Alpha explicitly logs prompts and completions to improve the model. The irony of “AI makes content free” is that these systems still rely on human intent and human-created data.
Third, distribution tactics are converging across modalities. For video, the winners appear to be platforms that can turn creation tools into more monetizable minutes. For text, the winners increasingly look like models that can win developer trust and platforms that can intermediate demand and route it.
With that…
Gong hei fat choi. Happy LNY! Unless something huge happens next week, I’ll be offline stuffing my face with yumz, and whether you celebrate or not, go enjoy some dumplings.
Some parting wisdom from our favorite immigrant tiger mom Jessica Huang here: A dumpling is always good, no matter how it looks. Link Here.
Gong hei fat choi.