DeepSeek launched its V4.1 Flash model on September 10, pricing it as low as $0.15 per million tokens during off-peak periods and highlighting sharply lower memory requirements. The model has 552 billion total parameters but activates 8 billion to 16 billion during inference. Its KV cache requirement is 890 bytes per token, one-quarter that of the previous V4 Flash, while SSD capacity requirements have fallen to one-eighth. The architecture divides 40 neural-network layers between an encoder that compresses documents and a decoder that generates responses, with multiple layers sharing cache. DeepSeek says V4.1 Flash outperforms Moonshot’s Kimi K3 on some benchmarks, although it remains behind leading Anthropic and OpenAI models overall. Micron Technology and SK Hynix ADRs each fell more than 5% at one point in New York, while Samsung Electronics and SK Hynix declined 2% to 3% in South Korea. MiniMax Group and Zhipu AI fell more than 8% in Hong Kong, and Alibaba dropped more than 2%. DeepSeek plans to phase out V4-Pro from September 14 and route inference workloads to V4.1 Flash, adding to concerns that lower-cost, more memory-efficient AI could pressure semiconductor demand and accelerate model commoditization.