DeepSeek V4.1 Flash takes on frontier AI models at a fraction of the cost

DeepSeek’s latest model was reportedly tested at 90% lower cost while pushing out more than 237 tokens per second, according to a widely shared post on X. The model in question is DeepSeek-V4.1-Flash, and it is quietly forcing an uncomfortable question onto every pricing team at the big AI labs: what exactly are customers paying extra for? What DeepSeek actually shipped DeepSeek launched V4.1-Flash on September 10, 2026. The model carries 552 billion total parameters. It uses a sparse mixture-of-experts design, or MoE, where only about 8 billion parameters are active when the model processes input and roughly 16 billion kick in when it generates output. It also ships with a 1 million token context window, allowing the model to hold a very large amount of text in working memory at once, such as a sprawling codebase or a stack of lengthy documents. Native image processing comes built in. The weights are open, released under the MIT license on Hugging Face, so developers can download and run the model themselves rather than renting access through an API. On October 17, 2026, DeepSeek pushed an update, labeled V4.1-Flash-1017, under the same API. The refresh improved performance metric...