DeepSeek Unveils V4 Preview: 1M Token Context MoE Models

DeepSeek released preview versions of its V4 model series on April 24, 2026, featuring DeepSeek-V4-Pro with 1.6 trillion total parameters and 49 billion active, and DeepSeek-V4-Flash with 284 billion total parameters and 13 billion active. Both support a standard 1 million token context window, trained on 33 trillion and 32 trillion tokens respectively, using innovations like token-wise compression, DeepSeek Sparse Attention, and hybrid attention mechanisms. These reduce inference FLOPs to 27% and KV cache to 10% of predecessor V3.2 levels for Pro at 1M context, enabling cost-effective long-context processing. The models offer thinking and non-thinking modes, with Pro-Max for high-reasoning tasks, and are available via API compatible with OpenAI and Anthropic formats, plus open-source weights on Hugging Face.