TMTPOST -- The latest artificial intelligence (AI) frenzy ignited by DeepSeek and other Chinese firms lifted valuation of TikTok owner ByteDance Ltd.

Credit：Doubao

ByteDance has been valued at more than $400 billion by at least three major investors, Bloomberg reported. Two of these investors, Fidelity Investments and T. Rowe Price Group Inc., have marked up ByteDance’s valuation to above $410 billion and $450 billion, respectively, according to Bloomberg calculations based on November filings to the US Securities and Exchange Commission (SEC). SoftBank Group Corp.’s Vision Fund, another investor, revalued ByteDance to north of $400 billion in December, factoring in its strong revenue growth last year and potential for its AI business Doubao, the report quoted people familiar with the matter.

The report suggested SoftBank may raise its estimated ByteDance valuation. For the Japanese company’s methodology assess TikTok’s U.S. operations as zero-value assets for the short video application was supposed to be shut down in the country.

The $400 billion valuation represents a sharp increase for ByteDance whose stock exchanged hands at a low point of $275 billion in July 2022. It is also a material increase compared with the value of $300 billion that set by ByteDance’s annual sales buyback last year, per the report.

ByteDance upgraded its flagship model large AI model Doubao on January 22, two days after DeepSeek released open-source DeepSeek-R1, the reasoning models that it claims performance comparable to leading offerings like OpenAI’s o1 at a fraction of the cost. ByteDance stressed the closed-source Doubao 1.5 Pro was trained by a "resource-efficient" approach that does not sacrifice performance. The model adopted an integrated train-inference design from the pre-training phase to balance between the best performance and most optimal inferencing cost, ByteDance said.

Doubao 1.5 Pro, with a Deep Thinking mode, scores higher than OpenAI o1 preview and o1 models on the AIME benchmark, and the model also outperforms DeepSeek-V3, GPT-4o and Llama 3.0-405B on popular benchmarks, according to Bytedance. Built on a new sparse Mixture of Experts (MoE) architecture,Doubao 1.5 Pro has activated parameters far fewer than the aforementioned models. It can deliver performance comparable to a dense model with one seventh the activated parameters of the later. Namely, 20 million activated parameters in Doubao 1.5 Pro equate to the performance of dense model with 140 billion parameters. It significantly exceeded the efficiency of conventional MoE architectures in the industry and achieves 7 times the performance leverage improvement.

Doubao 1.5 Pro integrates a heterogeneous system design for prefill-decode and attention-FFN tasks, optimizing throughput and minimizing latency. With extended context windows, the model supports 32,000 to 256,000 tokens for processing long-form text more effectively. It costs much less than competitor models, with operational cost 5 times cheaper than DeepSeek and more than 200 times lower than OpenAI’s o1. This breakthrough has been accepted by ICLR 2025, a top-tier AI conference, offering new insights into improving AI inference efficiency and scalability.

While domesitc peers Tencent Holdings Ltd. and Baidu, Inc. accelerate AI integration, ByteDance has reflected on its slower response to industry trends. ByteDance was slower than artificial intelligence startups to recognize and capitalize on the new opportunities in large language models (LLMs), CEO Liang Rubo said at an all-hands meeting on February 13.

Leading AI startups were set up and began work between 2018 and 2021, while ByteDance only started discussing frontier models such as OpenAI's GPT in its semi-annual technical review in 2023, Liang told employees. Liang noted ByteDance recognized the significant technological shift to the long-chain thinking model, which was one of features of DeepSeek’s reasoning model R1, but didn’t respond quickly enough.

ByteDance will focus on pursuing the upper limits of “intelligence” this year, rather than chasing the daily active users of any specific product, such as Doubao AI assistant, Liang said. It is hoped that the focus, putting the intelligence the foremost important, will inspire more experimentation and ensure critical technological milestones are not overlooked, according to Liang. That will include exploring new interactions and enhancing economies of scale, he said.