During Lunar New Year last year, China's state-level artificial intelligence (AI) company DeepSeek, which triggered the so-called "DeepSeek shock," was reported to be preparing to unveil the next-generation large language model (LLM) "V4." With the release likely around the Lunar New Year holiday next week, the next-generation LLM is expected to accelerate the innovation cycle of China's AI industry by focusing on lowering training and inference expense rather than competing on performance.
On the 12th, Chinese business outlet Huarjie Jianwen stated accordingly, citing Nomura Securities' "Global AI Trend Tracker" report. According to the report, V4 is expected to be unveiled around the 17th of this month, to capture Lunar New Year promotional effects as it did last year. However, the V4 to be released this year is likely to be closer to a structurally improved model focused on reducing training and inference expense, rather than a model that shocks the global AI market as last year's did.
For DeepSeek's most recent model, V3.2, the price per 1 million tokens is $0.28 (about 400 won) for input and $0.42 (about 600 won) for output. Compared with GPT-5.2 ($1.75 input, $14 output) and Gemini 3 Pro ($2 input, $12 output), it is up to dozens of times cheaper.
It is also expected to be competitive in terms of performance. Huarjie Jianwen said, "Internal tests indicated that V4's programming capability surpassed Anthropic's Claude and OpenAI's ChatGPT." Back in December last year, when V3.2 was released, some features were also evaluated as surpassing ChatGPT-5 and Google Gemini 3 Pro.
The report noted that, given U.S. semiconductor export controls have made it difficult for China to secure cutting-edge AI chips, V4 compensates for hardware limitations through algorithm and engineering optimization rather than hardware expansion. DeepSeek developed technologies called "mHC" and "Engram" and applied them to V4; put simply, mHC allows stable training even without top-tier chips, and Engram creates a structure that uses less memory. DeepSeek fused the two to pursue a strategy of "using current chips more efficiently" instead of "securing better chips."
The biggest effect of this kind of structural innovation is reducing training and inference expense. Until now, LLM and AI application developers have spent massive capital on training and inference; if this expense falls, demand will be stimulated, more companies will adopt AI models, the pace of commercializing services will quicken, and it will help improve profitability. This means China's AI industry could shift from a technology demonstration phase to a full-fledged revenue generation phase.
The market environment is changing as well. Unlike when DeepSeek's V3 and R1 once posted high shares in the open-source ecosystem, competing models are now rapidly multiplying, leading to a more multipolar market. Nomura Securities said, "It is difficult to dominate the market on the efficiency of a single model alone," adding, "The core of competitiveness going forward will be the ability to create differentiated services using the models, rather than the models themselves."