DeepSeek, a Chinese AI company, launched and opened sources for DeepSeek-V3 model on 26 December 2024, which outperformed similar open-source models in a number of tests, and was comparable to the top closed-source models in important areas, with low training costs, according to a research report issued by Guosen Securities. The model layer adopts the MoE framework, and outperforms open-source models in knowledge, code, mathematical reasoning and other tests after multi-stage training and capability refinement. Related NewsCICC Hikes MEITU's TP to $6.3; Cost Control Pushes Up 2025 Adj. Earnings ForecastThe architecture layer follows the V2 architecture and introduces new technologies, such as unaided loss-free balancing strategy and MTP to improve data utilization. The training layer realizes cost control and efficiency improvement through DualPipe algorithm and FP8 hybrid-precision training processor.