DeepSeek-V4-Flash API Launches; Xiaomi Raises Prices; LPDDR6 Nears Production
DeepSeek has launched the public beta API for DeepSeek-V4-Flash, which reportedly reduces single-task AI costs by 60% compared to GPT-5.6 Luna. Additionally, Xiaomi announced price increases for several phone series, and ChangXin Storage is nearing the production validation phase for its LPDDR6 memory. The release of highly cost-effective MoE models like DeepSeek-V4-Flash intensifies price competition in the LLM market, while the rise of microelectronics as China's highest-paying major reflects the country's strategic shift toward semiconductor self-reliance. DeepSeek-V4-Flash is an efficiency-focused Mixture-of-Experts (MoE) model featuring 284 billion total parameters (13 billion activated) and supporting a 1-million-token context window. Meanwhile, ChangXin's upcoming LPDDR6 memory boasts a speed of 12,800 Mbps with significant low-power optimizations over LPDDR5X.
## BACKGROUND
Mixture-of-Experts (MoE) is a machine learning technique where only a subset of the model's parameters (activated parameters) are used for any given input, significantly reducing computational cost and latency. LPDDR6 is the next-generation low-power memory standard designed to handle intensive AI workloads on mobile devices and servers.