According to Svmuu, DeepSeek has released a preview version of its V4 series open-source models under the MIT license, with model weights now available on Hugging Face and ModelScope.
This series includes two MoE models: the V4-Pro features approximately 1.6 trillion total parameters with 49 billion activated per token, while the V4-Flash has 284 billion total parameters and 13 billion activated per token. Both support a 1 million token context window. The company stated that, compared to the V3.2 version, memory usage and computational costs for long-text inference are significantly reduced.