As agentic AI pushes rivals to raise prices and cap usage, Deepseek ships a good-enough model for almost nothing

Deepseek launched V4-Pro and V4-Flash, its first new architecture since V3, and released them as MIT-licensed open weights on Hugging Face.

The models use a hybrid attention design to cut compute and memory for long contexts, while supporting up to a one-million-token context window.

Deepseek says the models are optimized for agentic workflows, trained on tens of trillions of tokens, and validated on Nvidia and Huawei hardware.

The company is pricing them well below comparable offerings from OpenAI, Google, and Anthropic, sharpening competition in high-capability AI.