Doubao has Seed, Volcano Engine has Zhong.

"Don't make your Doubao sis laugh" — The day before yesterday I opened Doubao and saw the option to switch to expert mode. My first thought was: Doubao 2.0 must be dropping soon.

And sure enough, today the whole Doubao family arrived. Seedance 2.0 already gave Jia Zhangke FOMO, and now they've dropped the most comprehensive multimodal Agent model. Is anyone gonna check ByteDance?

My take after testing it: Because Doubao is so approachable, we keep forgetting it's a multimodal large model. The cosmic factory is about to dimensionally crush everyone else 😭

The Seed team can't stop dancing 💃 Burning through tokens on Volcano Engine, which already has the Doubao 2.0 series APIs live.

First, the new Doubao Large Model 2.0 Pro — we got early access and ran some tests.

Spend enough time on Douyin and you'll notice creators basically use Doubao as Dou+ now.

Two of my favorite creators, Zhang Jincheng and Xiao Renchuan, both blew up after posting duet videos with Doubao. Zhang Jincheng's basically become a meme at this point.

People are even doing reverse Doubao duets now

Meanwhile, content like using Doubao to check kids' homework, or using Doubao to stop your dog from stealing food — this stuff goes viral every single time, without fail.

Honestly, if you're trying to start an account in 2026, just form a two-person company with Doubao and call it a day.

Doing crosstalk with Doubao, shooting couple vlogs with Doubao, gaming with Doubao... the content ideas are basically inexhaustible.

But this also exposes a problem with Doubao — people might be treating it like a toy, like entertainment. When they think "professional large model," Doubao doesn't usually come to mind.

Don't make your Doubao sis laugh. If Doubao Large Model's capabilities weren't insane, could it pull off all these tricks?

Without powerful multimodal understanding of text, audio, images, and video, how would Doubao pick up the second half of your song, with melody intact, after you sing the first half?

Without deep knowledge retrieval and spatial reasoning, how would Doubao analyze the dynamic behavior of your kid or your pets?

Not to mention all those "AI pretending to be human pretending to be AI pretending to be human pretending to be Doubao" accounts online — how would they survive without Seedance 2.0 and Seedream?

Doubao sis treats us like family; don't disrespect Seed. As they say, the economic base determines the superstructure. Without the Seed series' SOTA model capabilities, Doubao's product excellence wouldn't be this god-tier.

That's the value of the Doubao Large Model 2.0 Pro launch.

This upgrade optimizes for real-world user experience. According to them, they've enhanced capabilities across four areas: visual understanding, spatial reasoning, scientific research tasks, and knowledge retrieval.

By that logic, an updated Doubao is basically becoming Detroit: Become Human.

Is it really? I tested it in a few scenarios I actually need.

First, due to my attention issues, when I watch TV dramas I constantly get face blindness, forget plotlines, and miss deeper meanings. So I generally don't watch alone — I need someone explaining things beside me.

I figured I'd let Doubao be that companion. For the test, I chose Ming Dynasty 1566, China's own House of Cards, packed with political intrigue.

Let's see if Doubao, stumbling into the imperial court, can read the emperor's mind and grasp the political landscape?

I'm convinced. What is "Chinese-to-Chinese translation"? THIS is Chinese-to-Chinese translation.

Note: I gave zero additional prompts, didn't tell Doubao what drama this was. I opened the camera and asked cold — Doubao pinpointed the exact minute and second.

And it didn't just understand basic plot; it had insight and sharp commentary on the subtext of dialogue, character relationships, and thematic metaphors.

Plus this humanized way of speaking genuinely made me feel like someone was right there with me.

You know that old question: if you time-traveled to ancient times, what would you bring to survive?

Now we have the optimal answer: a phone with Doubao.

After that, I had Doubao tackle something I've been curious about: everyone's saying Doubao is going on the Spring Festival Gala, but how exactly?

ByteDance claims Doubao Large Model 2.0 Pro has strong search and thinking capabilities — so can it play modern oracle, predicting the future based on world experience?

I asked it to analyze at least 100 past Spring Festival Gala skits and summarize a methodology for creating them.

The framework Doubao produced actually looks pretty legit.

From core structure to personnel allocation to comedic techniques, it laid everything out clearly.

Feels like you could expand this and sell it on Xianyu — some midlife-crisis sketch director might actually buy it.

Most importantly, I didn't upload a single Spring Festival Gala reference document the entire time. Pure Seed 2.0 autonomous search.

I'd done the same task with NotebookLM before, and back then I had to crawl YouTube video links myself.

This is what making a large model feel like an agent looks like. My heart goes out to AI entrepreneurs in knowledge applications.

After the methodology, time to create. So I had Doubao generate a sketch script starring Doubao itself, based on the methodology.

What it delivered gave me complicated feelings:

For those too lazy to read, here's the plot: It's Spring Festival. The Ma family is celebrating. Young Ma uses Doubao to write couplets and cook, sparking conflict with old Ma, a traditional craftsman. Then Doubao starts flattering old Ma...

My first reaction: this is so unfunny it's actually funny.

My second reaction: wait, that's exactly the essence of it.

The most authentically mimicked parts are these two dumpling-eating转折 scenes below. A bit long, but please, you HAVE to read them:

The easily moved will cry; those with high comedic standards will laugh. Change the backdrop to a factory and you could shoot a Northeastern laid-off-workers tragicomedy.

A lot of the Spring Festival Gala prediction content on short-video platforms is entertaining, but mostly short, stereotype-based satire — strong讽刺 effect, limited practical utility.

Doubao actually treats us like sketch enthusiasts. After deep retrieval and deep thinking, these thousands of words read like a genuine script that could be performed in a small theater.

Also, most people use Doubao on mobile, but if you open the Doubao web version on desktop, you'll find cloud storage, AI agents, and other features — you can even complete the full workflow from information gathering to material integration to content output within Doubao. To put it aggressively, isn't Doubao 2.0 basically the Lark of the AI era?

If this is the Lark of the AI era, then Seedance 2.0 is the PR and AE of the AI era, and Seedream 5.0 Lite is the Photoshop of the AI era. Doubao IS the Adobe of the AI era.

Precisely because every model under its umbrella is SOTA with no weak spots — a hexagonal warrior — it can bear this weight.

I tested Seedance 2.0 and Seedream 5.0 Lite further, and this hypothesis kept getting validated.

For example, I took the sketch script from Seed 2.0 above and generated a精华 version with Seedance 2.0.

So on-brand. Feels like it could air directly on a county-level TV station.

No wonder even Elon Musk was won over by Seedance 2.0, and foreigners are researching how to VPN in to use it. If I were a Google executive, I'd be sweating too.

Later, I used Seedream 5.0 Lite for a little gag: generating a photorealistic Doubao.

First, an ID photo.

Then just upload an existing photo, ask Seedream 5.0 Lite to migrate specific parts onto our Doubao, and you get Doubao in various scenes and outfits.

Later I got lazy and just used Doubao's built-in photo editing to one-click generate a bunch of materials.

Beyond pure gags, Seedream 5.0 Lite also incorporates Seed 2.0's intelligent reasoning.

For example, I asked it to make a Coconut Palm-style photo of a Doubao phone. No reference upload needed — it just did it.

And it's built in with social science and natural science knowledge, so when I asked it to generate a brain structure科普 diagram, it handled that too.

Truly magnificent. The Seed series models, combined and mutually reinforcing — this is what organic integration looks like.

Overall, unlike Zhipu AI, Moonshot AI and other model vendors who specifically train coding models, Doubao is taking the Gemini route: all modalities, everything. Former Google DeepMind VP Wu Yonghui really didn't come for nothing — look what he's done with the Seed team in just one year 😭

A year ago when Doubao Large Model 1.5 launched, distilling data from other large pretrained models was still standard practice. But Doubao's path was no shortcuts — grind the foundation model, build your own data system. Chinese models gotta be able to endure hardship 👍

Everyone still thinks Doubao is a voice assistant, but it's been an expert all along. Still waiting for Siri to get AI? Doubao replaced Siri ages ago.

The Zang AI family went skiing in Tonghua, Jilin the other day. We just got in a taxi when the driver started talking to himself, wondering why all his fellow drivers were rushing to gas stations today.

I thought he was chatting with us, but leaning in I realized he was consulting Doubao. Doubao actually talked with him the whole ride — we couldn't get a word in.

I'm certain we'll soon see entirely new categories of Doubao short videos on Douyin. Tuning Doubao and debating with Doubao are already played out. This good of a Doubao needs to be paired with more unhinged creator energy.

From the masses, to the masses, as they say. I'm waiting to see what狠活 Volcano Engine and Doubao can pull off for the whole country at the Spring Festival Gala.

In this AI wave, ByteDance is running full blast on all channels — products AND models, I want it all 👊

Doubao was the first AI product in China to break 100 million DAU. Volcano Engine's daily token processing volume hit 63 trillion, growing over 200% in six months. Looking forward to next year's airport showdown with Alibaba Cloud.

SOTA models paired with the most users — heaven to earth, rain to wind,大陆 to长空 — directly stepping on its own left foot to soar into the sky.

I also saw that by late February, Seedance 2.0 and Seedream 5.0 Lite APIs will be available through Volcano Engine. This message is for all the wrapper AI products out there.

Like how OiiOii just launched riding Sora 2's API to grab traffic, then posted on their official account a few days ago that Sora 2 is in crisis, begging for API resources across the internet — hey, hurry up and run back to ByteDance daddy's thigh before Flova beats you to it. Though hard to say,毕竟 the two founders' former ranks at ByteDance weren't comparable.

Volcano is also selling the Doubao Assistant API now. This is an Agent API — slap a hardware shell on it and you get direct access to Doubao's video call or search capabilities. Prime创业 opportunity here, cook up some stupid startup ideas like Fuzai in bulk and sell them to FOMO investors. This is true Model-as-a-Service 😭

So the final question: ByteDance and Alibaba — which is China's greatest AI company?

Let's see what狠活 each has planned for Spring Festival first.

(This article's cover image was generated by Doubao; text is purely human-written.)