DeepSeek V4 arrives with two new models that offer high performance in long contexts at a low cost

The Chinese company DeepSeek has launched its new series of open artificial intelligence models, DeepSeek V4, con agent capabilities y improvements in the reasoning, in line with what current closed models offer, and offer high performance in large contexts at a reduced cost.

The DeepSeek V4 series, currently in preview, is composed of two Mixture of Experts (MoE) language models that integrate innovations in architecture and optimization that make them “highly efficient in large context scenarios”, with “drastically reduced” computing and memory costs“.

On the one hand it is DeepSeek V4 Pro, con 1.6 billion parameters and 49 billion active parameterswhich supports a context window of one million tokens. The Chinese company has highlighted the agentic capabilities of this model, especially in programming, which it already uses internally, as explained in X.

🚀 DeepSeek-V4 Preview is officially live & open-sourced! Welcome to the era of cost-effective 1M context length.

🔹 DeepSeek-V4-Pro: 1.6T total / 49B active params. Performance rivaling the world’s top closed-source models.
🔹 DeepSeek-V4-Flash: 284B total / 13B active params.… pic.twitter.com/n1AgwMIymu

— DeepSeek (@deepseek_ai) April 24, 2026

DeepSeek V4 Pro has also proven to outperform current open models in global awareness, being surpassed by the closed Gemini 3.1 Pro model. In reasoning, it rivals the main closed models.

On the other hand, DeepSeek V4 Flash has 284 billion parameters and 13 billion active parametersand a context window of a million parameters. It has been designed to offer faster responses without affecting its performance. DeepSeek claims that its reasoning capabilities are close to those of V4 Pro.

The company has launched these two models in a mode of early access in Hugging Face and can be tested both in the DeepSeek chat and in the API.