What to Watch in ByteDance's Doubao Models

Jiang Wu · 2024-05-15

*Originally published in Chinese on May 15, 2024.*

ByteDance, the owner of Douyin and TikTok, has launched its Doubao models. It went straight to vertical models and tied model capability to scenarios. This approach has several advantages.

Why scenario-specific models help

A lower barrier for developers

First, the barrier to use is as low as it can get. For many developers, being able to use an API well and being able to use it well inside a scenario are different requirements. To meet its own product needs, a team sometimes has to lay down basic functions first and add many details in between to keep the product's output stable.

More precise data

Second, the data is more precise. Developers wrap model capability directly into the product process, which requires a definition of the scenario capability, or they need the API to perform better in that scenario. So most people fine-tune for a scenario. Once Doubao models are split vertically by scenario capability, the barrier for developers is lower and the results are better.

Hardware reserves

Take the volume of text and images ByteDance says it processes every day. If we assume the enterprise users are negligible, then after subtracting a few other kinds of calls, you can almost work out its real activity today. The premise is that these users are all on the platform.

If it generates 30 million images a day and splits its core scenarios finely, that figure is formidable. In processing capacity alone, 30 million images a day is a heavy load for any cloud provider, so ByteDance should have a large reserve for this demand.

API pricing as bait

In logic, other model makers of the same type could also price this low. Lower prices are not very noticeable to consumers. For developers, they mean model capability can be used freely. Before, prices were set with reference to the model's price, so service fees were higher. If API prices are lower today and the scenario is more vertical, the more the front end charges, the larger the gross margin in between.

That depends on upstream and downstream API service prices, such as storage and distribution, staying stable enough to leave a profit as API prices keep falling. (The marinade is free, but customers cannot live on marinade alone.)

Doubao's numbers look conservative

I am not sure whether the published figures include CiCi, the overseas version. My sense is they do not. Across all versions, Doubao's daily active users should now be above 3 million. ByteDance also released products of its own today that use similar vertical models, and it looks like it is trying every dimension.

Scenarios and user feedback

Beyond these elements, what deserves the most attention?

ByteDance does not stress parameters, data or corpus. It split model capability vertically by scenario. I think the deciding factor is the difference between having applications and having none, and more fundamentally between having data and having none. With user feedback and data feedback, it can build more precise scenarios and more precise services.

When big companies face model scenarios and opportunities, they start from their own business needs, add model capability, turn it into a general product and release it outward. The first step is always an upgrade of their own services. A company that only makes models will face growing pressure from less and less user data feedback, and less direct scenario feedback. No obvious difference will appear in the short run. After a period without data, though, the gap in model capability will become clear, and the pressure will grow.

In short, big companies use model capability to serve their own products and users, and small companies have to find their own possibilities outside that large system.

So is the battle over models (a battle, not a war) finished?

Home · All articles · Topics · 简体中文