DeepSeek's New Efficiency Model Outperforms Rivals as Anthropic Sounds Alarm on AI Safety

DeepSeek released its V4.1 Flash model, claiming it outperforms previous flagships while drastically cutting costs—a strategic move as the Chinese AI lab prepares for its Shanghai STAR Market IPO. The model uses a Causal-Encoder-Decoder architecture and a Mixture-of-Experts design that activates only 8 billion parameters for inputs and 16 billion for generation, down from the full 552 billion parameter framework. This efficiency allows it to score 90.6...