Contrary to what many might think, artificial intelligence has neither faded away like the metaverse nor continued the frenzy of early-year breakouts. Like a continuously evolving digital wave, it has become an inseparable part of the industrial landscape.

I'd say this feeling is actually quite nice. After all, it's clear to see the practicality of the current batch of AI tools. I even mess around with these things from time to time. Believe it or not, with a bit of fiddling, creating text RPGs, novel illustrations, and even work tasks have become much easier.

As for recent news on domestic large models...

Oh right, ByteDance has new moves. The '2025 Volcano Engine Spring FORCE Conference' is here.

(Image source: Volcano Engine)

As a semi-annual conference, Volcano Engine did bring some noteworthy updates this time. Besides the comprehensive refresh of the Doubao Large Model family, there were also highly anticipated news on Kouzi and TRAE, along with immersive exhibition areas and over 10 special forums.

Want to know what new things Volcano Engine is up to next? Just follow me.

Doubao 1.6, here it comes!

Among domestic large models, Doubao's start was relatively slow.

But perhaps due to technological accumulation and late-blooming, the Doubao AI ecosystem achieved rapid development in 2024 and is now catching up.

According to statistics from research institution @QiyiFactor, ByteDance has now become one of the technology companies with the most comprehensive generative AI models and AI applications. Doubao's MAU is far ahead, seemingly ready to pull other similar Chinese chatbots into one group.

So, what surprises did Doubao bring us this time?

First, the highly anticipated underlying large model, Doubao Large Model 1.6, is officially released!

(Image source: Leitech)

According to Volcano Engine President Tan Dai, Doubao Large Model 1.6 includes three large models: Doubao-Seed-1.6-thinking, Doubao-Seed-1.6, and Doubao-Seed-1.6-flash, all supporting multimodal input and achieving 256K ultra-long contexts.

Taking the thinking model as an example, its thinking ability has been enhanced. Facing the new national math exam for college entrance, Doubao Large Model scored 144 points, ranking first among large models nationwide. In the Haidian mock exam, it scored 706 points in science and 712 points in liberal arts, both significantly higher than its predecessor.

The all-new Doubao-Seed-1.6 is somewhat like Gemini-2.5-Flash, supporting three thinking modes: on, off, and auto, allowing users to choose based on usage scenarios or letting the large model decide whether to use deep thinking, thus reducing token consumption while ensuring experience.

(Image source: Leitech)

Currently, these capabilities have been launched on the Large Model Application Lab, facilitating interested individuals and enterprises to develop their own agents.

Then, as previewed, Doubao officially released the video generation model Seedance 1.0 pro at this conference.

(Image source: Leitech)

In the field of video creation, Volcano Engine and its parent company ByteDance are perhaps the most authoritative.

The global video creation craze sparked by Douyin has ushered in a new era of internet videos. As such, Seedance 1.0 pro places more emphasis on users' creation processes and effects in practical use, bringing an underlying logic innovation for multi-camera narrative.

For example, users of traditional video generation models should know that AI-generated videos often suffer from 'jumping' due to spatial logic breaks during camera transitions. It's best to fix content within a single shot, generate in segments, and manually splice them together.

Seedance 1.0 pro, however, uses multimodal positional encoding technology to convert spatial information in text instructions into motion trajectories in a three-dimensional coordinate system, thus smoothly handling the relationships between characters, scenes, and cameras. Judging from the samples exhibited on-site, the overall transition is smooth, with minimal discomfort.

Oh, and there's big news: the Doubao real-time speech large model is also fully launched today.

(Image source: Leitech)

In fact, whether it's the previously popular AI video calls or the recently launched AI podcast generation, this large model's capabilities have been utilized, and the final results are indeed stunning. ByteDance naturally hopes more developers and users can leverage this capability.

Amid the wave of AI interaction evolving from text to multimodal, the launch of Doubao's series of new models seems to signify that the era of AI that can 'listen, see, and think' is quietly approaching.

Thriving Ecosystem

'Security is the foundation of all agents.' said Volcano Engine President Tan Dai.

Addressing the concerns of many enterprises that 'the cloud is not secure and on-premises is not user-friendly,' Volcano Engine officially released two enterprise-grade new technical solutions at today's conference: AICC Confidential Computing and Large Model Firewall.

(Image source: Leitech)

The former is easy to understand; it's what we often call 'data is usable but not visible.' Enterprises provide data to organization two through a confidential environment. In this environment, organization two can use this data but cannot see it, thus addressing enterprises' concerns about data leakage.

The latter provides security protection for enterprise-grade large model users. According to the official introduction, it can effectively resist computational DDoS attacks, reduce the risk of malicious token consumption by 30%, reduce sensitive data leakage by 70%, and control the rate of malicious information output within 5%.

How to reassure enterprises in using cloud AI services has become a common topic among service providers this year.

TRAE, which many domestic developers are concerned about, also achieved breakthroughs in three core capabilities at this conference.

(Image source: Leitech)

At the context understanding level, Trae has advanced from simply recognizing text content to deeply analyzing users' creative intentions. The MCP module endows AI with execution capabilities, enabling it to invoke external tools and services, akin to having 'operational hands.' The agent system constructs an 'expert advisor' mode, supporting flexible customization of workflows for different tasks.

The synergistic operation of the above functions, especially the deep integration with databases like MySQL, is expected to open a new chapter of efficiency leap for developers, significantly enhancing the convenience and efficiency of development work.

As an AI IDE launched by ByteDance, Trae is indeed a great inclusive tool that allows more people to master coding technology. With the introduction of these new features, it has transcended the category of 'AI editor' to become a partner that can fight alongside you.

According to Hong Dingkun, over 80% of engineers within ByteDance use TRAE to assist in development.

By the way, taking advantage of the break after the main forum, I also visited the exhibition area.

One of the more interesting things was a viewing area built based on videos generated by Seedance 1.0 pro. Don't get me wrong, the animation effect is quite impressive, but the heavy use of close-ups clearly doesn't align with the objective laws of real animation – no animator would be willing to manually keyframe so many details.

(Image source: Leitech)

Another product, this time ByteDance and its partners brought a plethora of products. Let's skip the more common ones like phones, cars, watches, and headphones. There were also wandering robots, AI toys with voice interaction, keyboards, mice, and a series of new AI categories.

(Image source: Leitech)

I'd say compared to the exhibition area half a year ago, this one is significantly more abundant.

Summary

Among domestic large model vendors, Volcano Engine's achievements are quite remarkable.

As of the end of May 2025, the daily average token invocation of Doubao Large Model has exceeded 16.4 trillion, ranking first among public cloud large model service invocations in China.

(Image source: Leitech)

Currently, Doubao Large Model has been widely deployed in industries such as automotive, smart terminals, internet, finance, education and research, retail and consumption, covering 400 million terminal devices like Xiaomi, OPPO, vivo, Honor, Lenovo, and Samsung; 80% of mainstream automakers, as well as dozens of securities and fund companies, numerous banks, top universities, and research institutes.

Watching the entire conference, what ByteDance/Volcano Large Model is doing is quite understandable:

'It must be user-friendly and cost-effective.'

This is also why Doubao 1.6 pioneered a unified pricing range. In the 0-32K input range, which is most used by enterprises, the input price of Doubao 1.6 is 0.8 yuan per million tokens and the output is 8 yuan per million tokens, with a combined cost only one-third that of Doubao 1.5's deep thinking model or DeepSeek R1.

(Image source: Leitech)

The same goes for the video model. The Seedance 1.0 pro model costs only 0.015 yuan per thousand tokens, and generating a 5-second 1080P video costs only 3.67 yuan, with 720P videos being even cheaper, at the lowest level in the industry.

Remember short videos in 2019? At that time, Douyin/Kuaishou were burning money madly, and many people called them foolish.

Now? Short videos have become the largest traffic entrance.

In my opinion, today's AI agents are like short videos back then. ByteDance can't wait because they know that whoever establishes an AI ecosystem first will gain the upper hand in the next era.

At today's conference, Volcano Engine upgraded its AI cloud-native full-stack services, releasing products like Volcano Engine MCP service PromptPilot intelligent prompting tool, AI knowledge management system, veRL reinforcement learning framework, and introducing a series of AI Infra suites.

All these contents are aimed at significantly lowering development thresholds, enhancing development efficiency, allowing developers to focus on creating truly good applications without worrying about model capabilities, costs, or development tools and platforms.

As the saying goes, groundbreaking new technologies and great products often start with uncertain experiments, finding a fit with market demand through iterations before growing and thriving.

Today, a drastic change is occurring in the cloud computing and AI services market, with Volcano Engine, a subsidiary of ByteDance, as the disruptor.

Behind Volcano Engine's disruption of the market landscape lies a comprehensive competition of resources, strategy, and execution. In an increasingly competitive market, it not only achieves a 'breakthrough' with low prices but also quickly establishes user awareness with capabilities like Doubao, Coze, and Trae, driving an overall leap in technological capabilities, ecological resources, and business models.

Perhaps this is Volcano Engine's winning strategy in the AI era's cloud war.