Zhang Yiming began to explain himself.

Zhang Yiming rarely explains himself.

But this time, he did.

According to reports from LatePost, two weeks ago, at an all-hands Seed meeting, he spoke continuously on several matters, with the core focusing on these keywords:

Don't be anxious, it's okay to fall behind temporarily, no distillation, Coding matters but don't only focus on Coding, AGI is the ultimate goal...

If this were just a technical discussion, these remarks would not carry much weight. But placed in the context of ByteDance's current position, it is more like the founder publicly reassuring the entire organization.

The last time Zhang Yiming was publicly described as delivering a speech at an internal meeting was a very long time ago. He has long been known for his low profile and in-depth thinking, and the ByteDance he leads is accustomed to letting results speak for themselves.

But today, the results have begun to blur.

I. ByteDance, Facing Public Scrutiny for the First Time

For a very long period of time, ByteDance itself was the best proof of strategic correctness.

This is not praise, it is a fact repeatedly verified by the market. From information feeds to short videos, from Douyin to TikTok, every memorable product launched by the company in the past decade was the first of its kind to break through in the market.

While others were chasing ByteDance, ByteDance only needed to prove it was right with its growth curve.

No one would ever ask "What on earth are we doing?" in an organization that has never fallen behind.

The situation has changed today.

The reception of the Seed 2.0 large model has been limited. GLM-5 and Kimi K2.5 have obvious leading advantages in Coding capabilities. In the Agent track, Tencent's WorkBuddy has pulled ahead. In the office scenario, Alibaba is integrating its Qwen office solutions. ByteDance is not in first place in any of these three directions.

Seedance is the exception. Seedance 2.0 is widely recognized as the world's highest-performance video model. But it precisely makes the lag in language models more conspicuous, proving that ByteDance has the capabilities, it just hasn't deployed them in the right places.

When businesses in different directions deliver completely opposite results, internal doubts about the route within the organization will surface.

From the release of Seed 2.0 in February this year to the launch of Claude Opus 5, GPT-5.6 Sol Max and even Kimi K3 today, ByteDance has fallen behind in its main model not by one month, but by a full half-year time window.

This is completely unprecedented in ByteDance's 10+ year history of never falling behind the curve.

When internal anxiety accumulates to a certain level, differences over the route will turn into huge internal friction.

Zhang Yiming stepping forward at this time usually means that such voices within the organization have become loud enough that the founder himself must respond and calibrate the direction.

II. Distillation Is Not the Real Problem

Almost all media outlets have focused the spotlight of this speech on Zhang Yiming's opposition to distillation.

But in fact, this is only one of the superficial manifestations.

What he is actually responding to is an increasingly impatient organization.

The Coding commercialization track was first seized by Anthropic, and ByteDance's overall profit growth rate has slowed down due to the drag of high computing power investment. The growth rate of Token consumption in the free Seedance scenario has declined, while paid enterprise invocation is still growing.

The daily average Token consumption of the Doubao large model climbed from 120 trillion in March to 180 trillion in June. It sounds like a considerable increase, but compared to the phase target of 250 trillion to 300 trillion set internally, the gap is actually widening.

Volcano Engine's total revenue for the whole year of 2025 is about 15 billion yuan. Its overall revenue target for 2026 is over 20 billion yuan, and its MaaS business target has been raised to 15 billion yuan. Its growth is highly dependent on video models, and there are structural hidden worries for long-term growth. Zhipu's ARR exceeded 1 billion U.S. dollars in July, while Kimi's ARR exceeded 300 million U.S. dollars in June.

All figures point to the same thing: Fast, fast, fast.

Thus a question naturally arises: if you want the model to run faster, why can't you use distillation?

Distillation is certainly not wrong.

But the problem is, it can only make you more like others.

From OpenAI to Anthropic, from Google to Meta, all leading players are using it. This practice should not be stigmatized, as it is essentially industrialization. Every mature industry will go through this path: Reverse Engineering, benchmarking, learning, replication, and then optimization. The automotive industry went through this, the chip industry went through this, and the internet industry is no exception.

Didn't Douyin learn from Instagram back then? Didn't Toutiao learn from existing information feed products?

They did.

The key is never whether you learn or not, but whether you have your own second step after learning.

For Zhipu, catching up with Claude after distillation is that second step. For Moonshot AI, K3 is one of the bargaining chips for its upcoming listing. But for ByteDance, this path is not enough. It only answers the question of how to catch up, not how to surpass.

Distillation can only let you approach an existing ceiling, it cannot help you build a higher ceiling.

Apart from the long-term strategic consideration, there is another dimension that is often overlooked, that is Zhang Yiming's strong personal technical purism.

As a tech enthusiast who advocates first principles, Zhang Yiming has an instinctive rejection of shortcuts that adopt ready-made results. Distillation essentially uses other people's thinking results to reshape its own model. For a founder who firmly believes in redefining the rules, it is not only a compromise on the technical path, but also a concession at the spiritual level.

His rejection of distillation is not only a strategic ban on laziness within the team, but also a projection of his personal geek purism on organizational decision-making.

Because he knows that once distillation is fully implemented in the organization, it will be very difficult to back out.

III. The Model No Longer Determines Everything

Another headache for ByteDance is that the industry has begun to evaluate AI with a completely different set of standards.

In the early days of large models, the core question was "whose model is the best?". ByteDance leveraged this to seize the entry point of AI assistants. By the first half of this year, the question suddenly changed to "whose work can be completed best?".

Tencent's WorkBuddy bet on the right track. It does not pursue how powerful the model is, but focuses on whether it can help people get work done. Alibaba's Qwen Office is also moving in this direction.

But ByteDance is still sticking to the Doubao route.

Liang Rubo, CEO of ByteDance, said at the all-hands meeting on August 6 that Doubao can serve as the "rough main line" to drive more businesses and ecosystems. This logic is the same as how Douyin drove e-commerce and local life services in the past. Douyin was the entry point, and everything started from Douyin. Doubao is the entry point, and everything starts from Doubao.

The problem is that these two entry points are not the same. Douyin solves the problem of what to watch, and the needs of all users are universal. Doubao solves the problem of how to get things done, and what is needed in different scenarios is completely different.

In previous articles, I mentioned that Zhang Yiming has poured over 50% of his energy into Seed recently, trying to reproduce the overwhelming technical suppression. But the reality is that Doubao's premature success, while bringing huge traffic to ByteDance, has also evolved into a strategic burden for the organization.

This inertia of the super entry point inevitably tilts ByteDance's resources and attention towards Doubao, leaving them no time to take into account other highly disruptive dedicated AI entry points.

When the team indulges in maintaining the main entry point, revolutionary opportunities in vertical scenarios are being carved up by competitors.

Doubao now has more than 300 million monthly active users, and ByteDance dares not let go, nor will it let go.

In ByteDance's history, the entry point is everything. Douyin's experience tells them: seize the entry point, and you can take your time with everything that comes after.

The problem is that today competitors are competing on work scenario Agents, while ByteDance is still competing on model capabilities.

These are competitions on two different dimensions.

IV. Seedance Proves One Thing

Since the industry evaluation system has shifted, why does Zhang Yiming still bet on the underlying model, even devoting half of his energy to Seed?

Obviously, he is betting that the evaluation system will eventually return to the model.

If all Agents become homogeneous in the future, the capability of the underlying model will ultimately determine the difference between players. Falling behind by half a year today is not a big deal. As long as the model catches up, the application layer on top will naturally be reshuffled.

Seedance has already proved this logic of counter-trend expansion.

In 2025, the mainstream view in the video generation industry was that further expanding the pre-training scale of the DiT architecture would yield limited returns, and the entire industry collectively shifted to a route dominated by fine-tuning and post-training. Kuaishou's Ling, constrained by its computing power budget, abandoned the large base model route and chose medium-sized models to focus on application implementation.

Seedance made a reverse decision, extending the pre-training stage to the extreme and increasing investment in computing power. The final product Seedance 2.0 has a parameter scale of 200 billion, and has fully implemented the MoE architecture, the first ever in the video generation field. After its launch in February this year, industry evaluations have uniformly ranked it at the very top of the global first tier.

According to exclusive news from LatePost, ByteDance is internally discussing training a large model with a total parameter scale of over 5 trillion. This scale surpasses Alibaba's Qwen 3.8-Max (2.4 trillion) and Moonshot AI's K3 (2.8 trillion), making it the largest planned model in publicly available information in China so far.

The project is led by Xiang Liang, head of the Seed Foundation, with Shen Ke, head of LLM pre-training data, coordinating the promotion. It is still in the early discussion stage, and no conclusion can be drawn on whether it will finally be implemented.

But the R&D logic of language models and video models are completely different. Seedance only needs more than ten core algorithm backbones to finalize the technical route and achieve breakthroughs. Training a 5-trillion-parameter language model requires undertaking a giant distributed system engineering of a completely different magnitude, processing a larger volume and more diverse full-domain text data, which requires in-depth collaboration of multiple departments across the company, and cannot follow the small-team breakthrough model used for video models.

This is not ByteDance's first bet. From Toutiao to Douyin, and then to TikTok, every key leap is a gamble. It's just that all previous gambles paid off, and in hindsight they look more like farsighted decisions.

This time the stakes are higher, and the result remains unknown.

V. Long-termism Has a Price for the First Time

Worse than the lag in technical indicators, what Zhang Yiming is truly wary of is that the entire organization has begun to subtly accept being the second place.

The capability boundary of a company never depends on what technologies it masters, but on what it believes it can redefine. Almost all of ByteDance's historically explosive growth products did not win by catching up with the first place, they all opened up a completely new game.

Algorithms become obsolete, talents leave, computing power and capital are consumed. Only the capability to redefine industry rules, once lost, can never be retrieved.

Decisions driven by anxiety will only make you chase in the direction where your competitors are running the fastest.

Over the past decade, ByteDance has always believed in one sentence: The best technology will eventually become the largest product. Recommendation algorithms, Douyin, and TikTok are all products of this rule. Long-termism has almost never failed to pay off.

But AI, for the first time, has pushed this redefining gene into an unprecedented commercial dilemma.

Today, the models that have achieved real commercial monetization are not necessarily those with the strongest underlying technology; the fastest-growing routes are not necessarily the ones that will eventually lead to AGI. Coding is not the end point, but it has become a commercial entry point first; Agent may not represent the upper limit of intelligence, but it has begun to reshape enterprise procurement and developer habits.

While competitors are quickly reaping gains by cutting into workflows and production systems, Zhang Yiming still devotes more than half of the company's resources and energy to the underlying breakthrough that may take years to deliver returns.

Thus, for the first time, the company has to answer a question: Are you chasing today, or chasing the future?

This leads to the most profound proposition in the AI era: being long-term correct may no longer equal being commercially correct.

In the past, long-termism only meant patience and determination. Today, it means extremely high opportunity costs. The longer you persist, the more certain short-term gains you give up.

What Zhang Yiming is doing this time is not sticking to a direction that he already knows will win, but in a desperate situation where the outcome is unclear and the opportunity cost is extremely high, he still chooses to reject the vulgar shortcuts and stay at the most difficult gaming table.

Words Beyond the Layout:

Zhang Yiming is persisting. But in fact, he is also fighting.

Fighting against the thing that is most likely to appear in an increasingly large company: the hidden short-termism.

Revenue, growth, user volume, all these things can justify themselves. Only basic research can't explain anything. It is invisible and intangible.

Every giant company will at some point start to wonder if it should speed up. Google, Apple, Tencent, Alibaba... have all gone through this moment.

Now it's ByteDance's turn.

The real problem has never been whether distillation is right or wrong.

It is whether a company that once won all wars with long-termism can still continue to believe in long-termism.

This article is from the WeChat official account "Beyond the Layout", author: Huahua, published with authorization from 36Kr.