China's DeepSeek launches V4 Flash and 1.6 trillion-parameter V4 Pro
DeepSeek, a major player in China's AI scene, just launched two new language models: V4 Flash and V4 Pro.
These models can process over 1 million tokens at once, so they're built for tackling massive documents or even entire codebases.
The V4 Pro stands out as the biggest open-weight model yet, with 1.6 trillion parameters, beating out rivals like Moonshot AI and MiniMax.
DeepSeek models near GPT-5.4 coding benchmarks
Both models focus on text-only tasks and use a mixture of experts setup, putting them close to OpenAI's GPT-5.4 on certain coding benchmarks, though they're a bit behind on general knowledge tests.
DeepSeek is also keeping things affordable: V4 Flash costs $0.14 per 1 million input tokens, while V4 Pro is priced lower than several named competitors, making high-powered AI more accessible for everyone.