DeepSeek V4 - almost on the frontier, a fraction of the price

This article covers the release of DeepSeek V4 Pro and V4 Flash, two new Mixture of Experts models from Chinese AI lab DeepSeek. The models feature 1 million token context, with Pro having 1.6T total parameters (49B active) and Flash 284B total (13B active). They are the largest open weights models available, using the MIT license. The article highlights their competitive performance against frontier models from OpenAI, Google, and Anthropic, while being significantly cheaper—Flash at $0.14/M input tokens and Pro at $1.74/M. DeepSeek's efficiency improvements, especially for long contexts, enable these low prices. The author also tests the models by generating SVG images of a pelican riding a bicycle and compares results with previous versions.