DeepThink, a Chinese artificial intelligence startup, is preparing to unveil its next-generation large language model (LLM) named 'V4' this month, creating anticipation regarding its ability to compete with the performance of models developed by leading U.S. companies like OpenAI and Google. Previously, DeepThink's model 'V3.2' launched in December last year had shown competitive performance levels against OpenAI's GPT-5 and Google's Gemini 3.0 in some evaluation metrics. The upcoming 'V4' model, featuring around 1 trillion parameters, is particularly noted for its superior performance in tasks requiring extensive coding prompts, which has garnered significant interest from the industry.

The model incorporates an advanced design known as 'manifold constraint hyperconnection' (mHC), allowing for more efficient data training while reducing reliance on costly high-bandwidth memory (HBM). Notably, analysts from Nomura Securities have highlighted that the inference cost per token of 'V4' could potentially decrease by 10-50% compared to its predecessor, making it an appealing option for businesses in need of cost-effective solutions. This could position DeepThink advantageously in a market that is rapidly evolving with new AI innovations.

As 'V4' approaches its launch, competition in the Chinese AI sector has intensified, with tech firms like ByteDance unveiling their AI chatbot Doubao 2.0 and video generation model 'Xidance 2.0' ahead of the Lunar New Year holiday. Additionally, Alibaba, a leader in the open-source segment, plans to release its own new model 'Q1'. The launch of V4 is crucial as it may determine DeepThink's standing in both domestic and international markets, especially against the backdrop of stringent U.S. semiconductor export regulations that have shaped the development pace of AI products in China. The outcome of this model's release could have far-reaching implications for the dynamics of global AI development.