China’s AI industry is no longer moving in one direction. DeepSeek and ByteDance, two of its most visible players, are showing how sharply the market is dividing between companies that want to build stronger models and companies that want to put AI directly into everyday software.
The contrast became clear when DeepSeek released DeepSeek V3.2 while ByteDance introduced more ways for people to use Doubao, its chatbot. Both companies have AI apps with over 140 million monthly users, but their latest moves point to very different ideas of what matters next.
DeepSeek is still betting on model performance
DeepSeek’s new release, DeepSeek V3.2, is an open-weight model that anyone can tinker with. The company says it performs on par with the latest models from OpenAI and Google and beats them on some key mathematics benchmarks.
The release did not fully satisfy every follower of the company. Some were still waiting for R2, the anticipated update to the initial model that rocked Silicon Valley in January. Instead, DeepSeek released V3.2 and V3.2-Speciale, described in the source as better-optimized versions of V3.2-Exp, which was released in September.
Even so, V3.2 drew attention because DeepSeek says it can solve the type of advanced math questions asked at the International Mathematical Olympiad. Its performance on other coding and reasoning tasks is also described as supposedly on par with or above GPT 5 and Gemini 3.
“It suddenly dawned on me why they call the company DeepSeek with the whale as a motif. Because just like a whale, it rarely surfaces, but every time it surfaces, it always makes a massive splash,” says Jen Zhu Scott, an AI investor and the cofounder and CEO of Power Dynamics, a modular data-center solutions firm.
DeepSeek’s position is not only about headline benchmarks. The company has also separated itself through efficiency. From the beginning, it has focused on model efficiency, partly because it reportedly does not have access to an abundant supply of computer chips.
That matters because efficient models can use fewer computing resources to train and cost developers less to run than those from Western labs. For startups and developers, that can make the difference between a model that is impressive in theory and one that is practical to use.
“Efficiency is very important if you are a startup. An open source model is superior both in terms of costs and because of customizability,” says Steve Hsu, a physics professor at Michigan State University and an AI startup founder.
ByteDance is moving Doubao closer to the phone itself
ByteDance is taking a different route. It is not making smartphones, but it is trying to make Doubao feel like a deeper part of how a smartphone works.
In late November, ByteDance released Doubao Input Method. Input apps, also called keyboard apps, help people type Chinese characters quickly on QWERTY keyboards. The source describes them as every Chinese person’s portal into the digital world.
Doubao Input Method gives ByteDance another entry point into its AI app ecosystem. The company says the keyboard app offers superior speech-to-text functionality, which places Doubao closer to the moment when a user starts typing, speaking or searching.
The more ambitious move came on Monday, when ByteDance released a Doubao AI agent that can be integrated into a smartphone’s operating system. The goal is to give the agent control over any app.
In a preview video, ByteDance showed Doubao using voice inputs to access Tesla’s app and open the trunk, search through different ecommerce platforms to find the lowest prices, and access photos in a user’s camera roll and enhance them with AI.
ByteDance is working with the Chinese smartphone manufacturer ZTE to preinstall the Doubao agent on one of its phone models, the Nubia M153. That phone sells for 3,499 RMB (about $500). ByteDance says it is also talking to other smartphone makers about installing its agent.
The challenge is that other major phone brands may have little reason to hand that position to ByteDance. The source notes that the most popular Chinese smartphone brands, like Huawei or Xiaomi, are developing their own proprietary AI agents.
WeChat shows the limits of an agent strategy
Doubao’s agent quickly ran into a major obstacle: Tencent’s WeChat. The app has over 1.4 billion users and an expansive range of features, making it as close to an operating system as China has today, according to the source.
On Wednesday, users on Chinese social media reported that they were suspended from WeChat after using Doubao’s agent to access it. The risk of losing access to WeChat is severe enough that few users would be willing to test it casually.
Doubao then announced that it had disabled its agent from working with WeChat and said the suspended users would get their accounts back soon. The incident shows that an AI agent can be technically ambitious and still be blocked by the platforms it needs to operate inside.
That matters for ByteDance’s wider strategy. If Doubao is meant to act across apps, then cooperation from those apps, or at least tolerance from their owners, becomes part of the product’s success. In China’s big-tech environment, that cannot be assumed.
Two models for where Chinese AI goes next
DeepSeek and ByteDance represent two ends of the spectrum. DeepSeek is still part of the model race, where companies compete through open source releases, reasoning ability, coding performance and mathematics benchmarks. ByteDance is focused on distribution, integration and repeat use.
Other Chinese AI companies fall between those poles. Zhipu, Minimax and Moonshot are following a path closer to DeepSeek by releasing capable open source models. Baidu and Tencent are more application-focused. Alibaba remains mostly on DeepSeek’s side through new versions of its open source Qwen models, though it recently released a consumer-facing AI super-app.
ByteDance has a reason to avoid the benchmark race. It already has a gigantic user base, and as a private company, Hsu says, it does not need to worry about stock market volatility.
“They just have to ship quietly, integrate a really powerful AI model into their existing app, and I think that’s what they are doing,” he explains.
The broader pattern is that Chinese AI firms are not simply copying the American tech giants’ approach. The source says they are steering clear of the compute hoarding game, where companies build more data centers and brute-force their way toward more powerful models.
Part of the reason is American chip sanctions, which block Chinese companies from accessing cutting-edge chips from Nvidia. Another reason is capital. As Hsu puts it, “they are just not as well capitalized. They don’t have infinite amounts of money the way that American companies have.”
That leaves Chinese AI companies with a practical choice. Some, like DeepSeek, are trying to make models more efficient and customizable. Others, like ByteDance, are trying to make AI more present in the daily flow of typing, searching, shopping, voice commands and phone use.
The split may matter more than any single release. DeepSeek is asking how powerful and efficient an open-weight model can become. ByteDance is asking how deeply an AI assistant can be woven into the software people already use. China’s AI race is now being fought on both fronts.