DeepSeek V4 in preview: two open models with one million token context

DeepSeek published on April 24, 2026 the preview version of DeepSeek-V4, its new generation of artificial intelligence models. The family comes in two variants: DeepSeek-V4-Pro, aimed at maximum performance, and DeepSeek-V4-Flash, designed for faster responses at a lower cost. Both are open source and can be downloaded from Hugging Face and ModelScope.

V4-Pro has 1.6 trillion parameters in total, but activates only 49 billion per query, allowing for efficient resource use. V4-Flash has 284 billion total parameters with 13 billion active per query. The size difference translates into speed and cost: Flash responds faster and consumes fewer resources, while Pro offers greater reasoning capacity.

The most notable feature of the family is the ability to process up to one million tokens of context, enabling work with long documents, extended conversations or large codebases without losing track. According to the company, V4-Pro requires only a fraction of the compute that its previous model V3.2 needed, thanks to improvements in how it manages attention across text elements.

According to benchmarks published by DeepSeek, V4-Pro outperforms all open-source models in mathematical reasoning, science and competitive programming, and comes close to the most advanced proprietary models. In general world knowledge, only Gemini-3.1-Pro ranks above it. Both models are also optimised to work as AI agents — executing tasks autonomously across multiple steps — with native compatibility with tools such as Claude Code.

The models are available at chat.deepseek.com and through the official app, with an API compatible with the OpenAI and Anthropic formats. The previous model identifiers deepseek-chat and deepseek-reasoner will be retired on July 24, 2026.

Key points

• DeepSeek launches DeepSeek-V4 in preview in two variants: Pro and Flash.
• Both are open source, available on Hugging Face and ModelScope.
• They support up to one million tokens of context.
• V4-Pro: 1.6 trillion total parameters, 49 billion active.
• V4-Flash: 284 billion total, 13 billion active.
• V4-Pro outperforms all open models in mathematics, science and coding.
• Only Gemini-3.1-Pro surpasses it in general knowledge.
• Optimised for AI agents, natively compatible with Claude Code.
• The API supports OpenAI and Anthropic formats.