DeepSeek Unveils V4 AI Model with 1‑Million‑Word Context and 1.6‑Trillion‑Parameter Pro Version

DeepSeek introduced its V4 artificial intelligence model, featuring a 1-million-word context capacity. The release includes Pro and Flash versions, with the Pro model containing 1.6 trillion parameters.

DeepSeek, a Chinese startup, previously disrupted the AI landscape with its low-cost R1 reasoning model. This earlier model matched US rivals' capabilities while requiring significantly less computing power.

This event, dubbed the "DeepSeek shock," influenced global discussions on AI strategy and innovation. The company's models have seen wide adoption within Chinese municipalities, healthcare institutions, and financial sectors.

DeepSeek-V4 now significantly expands this capability with an ultra-long context capacity of one million words. This allows the model to process substantially more input for complex tasks, enhancing its ability to understand and generate extended content.

The model comes in two versions: DeepSeek-V4-Pro and DeepSeek-V4-Flash. DeepSeek-V4-Pro contains 1.6 trillion parameters, which refine the model’s decision-making ability, while DeepSeek-V4-Flash features 284 billion parameters, offering a more efficient and economical choice.

In world knowledge benchmarks, DeepSeek-V4-Pro outperforms other open-source models. It trails only slightly behind Google’s Gemini-Pro-3.1, a top-tier closed-source model. The model also optimizes for popular AI Agent products such as Claude Code and OpenCode.

DeepSeek's strategy of open-sourcing its models contrasts with many Western proprietary systems. This approach continues to drive its adoption and influence the global open-source AI ecosystem.

This release further intensifies the global AI race, prompting increased focus on protecting national innovation in artificial intelligence. Observers will now monitor DeepSeek-V4's real-world adoption and its broader impact on competitive AI development.