ByteDance launches Doubao-1.5 Pro, maintains positive gross margins despite ...

by Xiao Fang

On January 22, ByteDance introduced Doubao-1.5 Pro, its latest closed-source multimodal model. The Beijing-based company emphasized its "resource-efficient" training methodology, which aims to optimize performance without sacrificing quality.

Doubao-1.5 Pro, reportedly outperforms GPT-4 in select benchmarks while significantly lowering operational costs. ByteDance credits the model’s large-scale sparse MoE architecture for this success, claiming it achieves performance equivalent to dense models with seven times the activation parameters, compared to the industry’s standard of threefold.

According to sources speaking to Jiemian News, ByteDance’s Doubao large language model continues to generate positive gross profit margins despite significant price cuts last year. The latest iteration, the Doubao-1.5 Pro model sold on the Volcano Engine platform maintains a gross margin of 50%.

When ByteDance launched the Doubao model in May 2024, it aggressively priced the pro-32k version at just 0.8 yuan per million tokens, a stark contrast to some competitors charging up to 120 yuan per million tokens. This strategy pressured rivals such as Alibaba Cloud’s Qwen to lower their prices. While questions about the sustainability of this pricing arose, industry analysts have since downplayed these concerns.

In a previous interview with Jiemian News, Tan Dai, President of Volcano Engine, ByteDance’s cloud computing services unit, highlighted the importance of high usage in refining models and reducing costs. “Lowering the price of large models isn't just about offering lightweight versions; the main and most advanced models also need to be affordable to meet the complex business needs of enterprises, validate the value of applications, and drive innovations that surpass current products and organizational models,” he said.

By late 2024, the daily token usage of Doubao models had surpassed 4 trillion—a 33-fold increase since its launch. Pricing for Doubao-1.5-pro-32k remains at 0.8 yuan per million tokens for input, 0.16 yuan for cache hits, and 2 yuan for output. The lighter Doubao-1.5-lite-32k is offered at 0.3 yuan, 0.06 yuan, and 0.6 yuan, respectively—making them among the most cost-competitive in their class.

Doubao’s applications have expanded rapidly, securing partnerships with 80% of major automotive brands and integrating into approximately 300 million devices, including smartphones and PCs. Usage across smart devices has increased 100-fold over the past six months. Despite these achievements, ByteDance’s large model business remains unprofitable due to heavy R&D expenditures. Continued growth in user adoption will be crucial for long-term profitability.