TikTok’s parent company, ByteDance, announced ‘Doubao 1.5 Pro’, an updated version of the popular Chinese AI app Doubao, on January 22, 2025. Chinese companies have limited access to advanced chips due to US export restrictions, but they are using efficiency to rapidly release AI that is low-cost yet has performance comparable to OpenAI’s models.
Doubao 1.5pro – Doubao Team
https://team.doubao.com/en/special/doubao_1_5_pro
China’s most popular AI app gets facelift with ByteDance’s Doubao 1.5 | South China Morning Post
https://www.scmp.com/tech/big-tech/article/3295834/chinas-most-popular-ai-app-gets-facelift-bytedances-doubao-15
DeepSeek claims its ‘reasoning’ model beats OpenAI’s o1 on certain benchmarks | TechCrunch
https://techcrunch.com/2025/01/20/deepseek-claims-its-reasoning-model-beats-openais-o1-on-certain-benchmarks/
Below are the benchmark results for the closed-source multi-modal model ‘Doubao 1.5 Pro’ announced by ByteDance.
Doubao 1.5 Pro outperformed OpenAI’s GPT4o, Google’s Gemini and Alibaba’s Qwen in half of the 14 benchmarks that measured coding skills, reasoning ability, Chinese language comprehension and more.
In its announcement, ByteDance said, ‘Doubao 1.5 Pro outperforms state-of-the-art ultra-large and dense pre-trained models with only smaller activation parameters, achieving outstanding results in multiple benchmarks,’ highlighting the company’s efficient training methodology, which combines performance with low cost.
China is steadily releasing models with performance that rivals leading AI models. For example, on January 20, 2025, the company announced the open source model ‘DeepSeek R1,’ which showed scores comparable to OpenAI o1 in several benchmarks.
Chinese AI company releases ‘DeepSeek R1’, an inference model equivalent to OpenAI o1, under the MIT license that allows commercial use and modification – GIGAZINE
According to a DeepSeek technical report, the model was trained on NVIDIA’s China-tuned H800 chip for less than $6 million, while Anthropic CEO Dario Amodei estimates that OpenAI spent $100 million to train GPT-4o.
TechCrunch, an IT news site, said of the rise of Chinese AI, ‘At least three Chinese research institutes — DeepSeek, Alibaba, and Moonshot AI’s Kimi — have developed models they claim are rivals to OpenAI’s o1.’
In addition, Dean Ball, an AI researcher at George Mason University, said in a social media post that he believes ‘Chinese labs will continue to be first followers in terms of achieving benchmark performance similar to that of American models.’
DeepSeek r1 takeaways for policy:
1. Chinese labs will likely continue to be fast followers in terms of reaching similar benchmark performance to US models.
2. The impressive performance of DeepSeek’s distilled models (smaller versions of r1) means that very capable reasoners…— Dean W. Ball (@deanwball) January 20, 2025