Zhang Yiming, faithful, expressive, and elegant.

(This article was written by Beyond the Layout and authorized for release by TMT Post)

Text | Beyond the Layout, Author | Hua Hua

In April 2026, DouBao reached 315 million monthly active users, ranking first among AI applications in China. In the battle for conversational AI on the consumer end, ByteDance emerged victorious.

However, in the same month, on another front, ByteDance seemed rather quiet. Spicy crayfish swept the nation, with long queues forming outside Tencent’s headquarters as staff helped users install software. Alibaba launched CoPaw, and companies were scrambling to capture the Coding Agent entry point.

What was ByteDance's move? Feishu introduced an OpenClaw plugin, while Volcano Engine followed up with ArkClaw. Functional, but far from an aggressive push. During the week when nearly a thousand people queued outside Tencent’s headquarters to help users install software, the main DouBao app still didn’t have a visible entry for crayfish.

Also that month, a personnel update shook the industry. Today, according to 'The Late Post,' former core researcher of DeepSeek, Daya Guo, joined ByteDance's Seed team, focusing on Agent technology. His total compensation package approaches the level of top domestic AI talent.

ByteDance later denied the claim of a 'near 100-million-yuan annual salary.' Li Liang, Vice President of Douyin Group, responded on social media, stating the report was inaccurate and emphasized that the Seed team has a unified salary system.

But this also indirectly confirmed the fact that Daya Guo joined ByteDance.

While the exact figures vary depending on who you ask, the signal is clear. ByteDance is stockpiling talent for a new war. And this war isn't happening inside chat boxes.

To understand why ByteDance is bringing in Daya Guo at this particular moment, one must first look at what Seed has been through in the past year.

ByteSeed is a large-scale foundational research team established in 2023. In February 2025, former Google Fellow Wu Yonghui joined ByteDance and took over Seed.

He worked at Google for 17 years, initially focusing on search ranking engineering before transitioning to Google Brain to drive the transformation of translation and search through deep learning, making him one of the highest-ranking technical experts within the company.

The situation he faces is challenging. At that time, this research team of over a thousand people had spent tens of billions catching up over nearly two years, finally training a foundational model ranked among the top tier domestically. However, in early 2025, their achievement was overshadowed by the results of a lean team from DeepSeek.

More troublesome is the issue of staff turnover. Multiple media sources confirm that in the past year, nearly 70 technical talents from the Seed team have resigned, moving to Tencent, Alibaba, large model ventures, and overseas tech giants. In the capital markets, more than 30 AI ventures under the ByteDance umbrella have completed financing. Seed has no shortage of talent reserves, but retaining them is a challenge.

Wu Yonghui's approach was to consolidate the team structure. According to public reports, after taking over, he held intensive discussions with hundreds of core researchers, reorganized the research direction, and unified internal data and code repositories. In July 2025, Zhou Chang took charge of multimodal interaction and world modeling, reporting directly to Wu Yonghui.

By the end of January this year, Liang Rubo was very direct during an all-staff meeting: By 2026, the AI model capabilities must rank among the industry leaders, and compensation and incentives should be ahead of top-tier levels across all global markets.

After a year, the foundation has stabilized. The base models are being iterated, multimodal systems are being monitored, and the training process is being streamlined. However, there remains one area lacking a leader who can take the helm — Coding and Agent.

Why has Coding suddenly become a track everyone wants to dominate?

The answer has already been written abroad.

Anthropic's Claude Code officially launched in May 2025. By February 2026, its annual recurring revenue (ARR) reached 25 billion US dollars.

During the same period, another benchmark product, Cursor, had just surpassed 20 billion US dollars in ARR, with its parent company Anysphere reaching a valuation close to 300 billion US dollars. However, Cursor took more than two years, while Claude Code achieved this in less than a year.

The significance of Claude Code goes far beyond its own revenue; it directly rewrote Anthropic's company destiny. At the end of 2025, Anthropic's total annual revenue was approximately 90 billion US dollars, while OpenAI's was 214 billion during the same period, with a gap of more than double. But by April 2026, Anthropic’s ARR skyrocketed to 300 billion US dollars, surpassing OpenAI's disclosed 250 billion US dollars in February. The growth driven by a coding product turned the underdog into the leader.

The overseas battle has become very clear. Coding is the main battlefield where Anthropic and OpenAI are fiercely competing, with Opus 4.6 against GPT-5.4 and Claude Code against Codex. Both models and products are simultaneously going head-to-head, even sparking public opinion battles at one point.

In China, the explosive popularity of Xiaolongxia pushed programming demands from programmers to the general public. However, big tech companies’ investment in the coding sector has been notably lagging.

According to LatePost, insiders close to ByteDance's relevant teams revealed that in 2025, every time they trained a Code model internally, 'higher-ups said it would be the last time.' Alibaba faced a similar situation, with Hui Bin, the head of Qwen Code, having already resigned.

The strategic emphasis on coding by leading tech giants only began after 2026.

Ventures, on the other hand, took the lead. Zhipu prioritized coding as a key optimization direction in early 2025, while MoonShot’s Kimi established Coding + Agent as the model’s core focus, continuously iterating. After Xiaolongxia triggered a surge in computing power demand, these ventures were the first to reap the benefits.

Zhipu CEO Zhang Peng revealed at the 2025 earnings briefing that API pricing increased by 83% in Q1 2026, yet usage volume still grew by 400%, with the market in short supply. MoonShot was even more impressive; according to media reports, within less than a month of Kimi K2.5’s release, its cumulative revenue over the past 20 days exceeded the total for all of 2025. Within two months, it completed two rounds of financing totaling over 12 billion US dollars, with its valuation doubling from 4.3 billion US dollars to over 10 billion US dollars.

On one side, overseas markets have already entered into fierce competition, while in China, Ventures have taken a half-step lead. The tech giants have just woken up.

Guo Daya, a post-95s individual. He is a Ph.D. jointly trained by Sun Yat-sen University and Microsoft Research, specializing in Code Intelligence. Google Scholar shows that his papers have been cited over 38,000 times, with an h-index of 33.

For a researcher who has been out of his Ph.D. for less than three years, this figure is quite rare.

After graduating with his Ph.D. in 2023, he joined DeepSeek. During his more than two years there, Guo Daya was the first author of the DeepSeek-Coder series and also a core contributor to star models such as DeepSeek V3, R1, and Math.

The GRPO (Group Relative Policy Optimization) algorithm proposed by him and his team in the DeepSeek-Math paper was later directly applied to the reinforcement learning training of R1. It was this algorithm that allowed R1 to spontaneously develop reasoning capabilities under limited computing power conditions, and it also made DeepSeek famous worldwide.

In other words, this person is both a core builder of Code models and deeply involved in the underlying methodology of reasoning models. In the context of Agent, these two capabilities almost cover the most fundamental technical foundations. For an Agent to function properly, the model must be able to understand code, generate code, and orchestrate complex tasks, which are precisely what Guo Daya excels at.

According to LatePost's report, Guo Daya had plans to leave as early as October 2025. One reason was his optimism about the agent direction, but at the time, the priority for agents within DeepSeek was not high. After leaving, Alibaba, Tencent, and ByteDance successively approached him.

It was reported that Alibaba offered him the position of post-train manager, but he ultimately chose ByteDance because their research direction was more aligned. ByteDance clearly prioritized optimizing the next-generation model’s Agent capabilities.

A noteworthy detail: even DeepSeek did not prioritize agents in 2025. This indicates that coding and agents have only recently started to gain significant attention from Chinese companies.

When several clues are viewed together, the outline of the track shift has become very clear.

The landscape for consumer-facing conversational AI is nearing convergence. DouBao has reached 315 million monthly active users, Qwen 203 million, and ChatGPT remains globally dominant; the experience gap among competitors is rapidly narrowing. What can be done within chat interfaces is becoming increasingly similar, leaving less room for differentiation.

High-value tracks have already shifted. Claude Code generated $2.5 billion in ARR in less than a year, Cursor surpassed $2 billion in just over two years, and Anthropic, driven by Coding growth, increased its annual ARR from $9 billion to $30 billion. These figures repeatedly confirm one fact: Coding represents the most proven and efficient commercialization path for AI today.

No one in this track in China has truly emerged yet. Ventures have taken the lead but remain limited in scale, while major firms have abundant resources but previously underinvested, and are now collectively catching up.

ByteDance has consumer-facing traffic, Feishu as an enterprise entry point, computing power, and the stable Seed foundation maintained by Wu Yonghui, and now they've added Guo Daya, someone tested through DeepSeek in both Code and reasoning directions.

But starting late is still starting late. Internally, Seed is undergoing organizational integration in the agent and Coding directions, and there's still distance between research and product. The window for ByteDance won't be open for long.

Zhipu has raised prices three times and is still in short supply, Dark Side of the Moon earned a year’s revenue in 20 days, Alibaba and Tencent are also aggressively recruiting talent—everyone sees the same opportunity.

The battle for conversational AI is nearing its end. The fight for Coding has just begun.

[Beyond the page]:

Faithfulness, expressiveness, and elegance were the standards set for translation more than a century ago by Yan Fu, a modern enlightenment thinker in China.

What ByteDance needs today aligns with these three words: technology must be trustworthy, business operations must be connected, and products must lead.

The name Guo Daya is a metaphor. But metaphors can't be eaten as food. What ByteDance truly acquired is not just an individual, but a ticket to the next big battle.

This battle is called Coding. It has already begun globally, and China has just entered the arena.

51K Views