DeepSeek releases V4, an AI model matching frontier performance at a fraction of training cost

DeepSeek (Chinese AI lab) released V4, a large language model (AI that understands and generates text) that performs on par with industry leaders like OpenAI's o1 and Claude 3.5, but was built at significantly lower computational and financial cost than Western competitors.

V4 uses a technique called mixture-of-experts (the model activates only the relevant parts of itself for each question, rather than using its full capacity every time), which means it answers queries faster while using less computing power — making it cheaper to run at scale than models that process everything through every layer.

For anyone building AI products or using AI services, this shifts the cost-performance equation: smaller companies and teams in cost-sensitive markets now have access to cutting-edge capability without enterprise budgets, while existing AI service providers face pressure to lower prices or improve their offerings to stay competitive.

DeepSeek publishes its models openly (free to download and use), making V4 immediately available to researchers, developers, and companies worldwide — no waitlist, no API key required.