TikTok owner ByteDance has surprised the artificial intelligence industry with the ultra-low cost of its Doubao model, which the company said is capable of processing 2 million Chinese characters, equivalent to 1.25 million tokens, for RMB 1 ($0.14). OpenAI’s most advanced multimodal model, GPT-4o, also unveiled this week, comes in at $5 per million input tokens handled.
“A good model works well and is affordable to everyone,” Tan Dai, president of Volcano Engine, ByteDance’s cloud computing services unit, said at an event on Wednesday. “Large usage polishes a good model and dramatically reduces the unit cost of model inference,” he added.
Why it matters: The radical pricing, which Tan explained is achieved through model structuring and hybrid scheduling of cloud computing power, is likely to initiate a price war in China’s AI field after a battle on parameters.
Details: ByteDance’s Doubao large model family consists of nine versions, ranging from a “lite version” to the advanced Doubao Pro, which is able to deal with up to 128,000 token inputs. The range also includes models centered on generating pictures and virtual characters.
- The company did not disclose how many parameters it used to train its most powerful Doubao Pro model.
- The Doubao LLM family shares the same name as ByteDance’s ChatGPT alternative, a free AI bot that the company says has 26 million monthly active users.
- The Doubao model is now open to external developers and other interested parties for cooperation via Volcano Engine, though this comes months after tech rivals Alibaba and Tencent gave enterprise customers access to their models last April and September, respectively.
Context: Earlier this year, ByteDance founder Zhang Yiming criticized employees for not being “sensitive enough” to the emergence of new technologies like ChatGPT in an internal meeting, saying that discussions among staff over ChatGPT only started in 2023. His remarks came after the company had already launched the Doubao chatbot in the second half of last year.