DeepSeek V4 Preview arrives with open-source models and 1-million-token context
DeepSeek has introduced V4 Preview with two variants, V4-Pro and V4-Flash, open weights, API availability, and a focus on long context and agentic capabilities.
Summary
DeepSeek announced `DeepSeek-V4 Preview`, a new generation of open-source models focused on long context, reasoning, and agentic tasks. The headline feature is support for a 1-million-token context window across the company's official services.
The release includes two variants: `DeepSeek-V4-Pro`, with 1.6 trillion total parameters and 49 billion active parameters, and `DeepSeek-V4-Flash`, with 284 billion total parameters and 13 billion active parameters. API access is already available through the `deepseek-v4-pro` and `deepseek-v4-flash` model names.
In practice
V4 Preview is especially relevant for developers, product teams, and companies that need to process large amounts of context. A 1-million-token window can support analysis of large codebases, long documents, project histories, contracts, text datasets, or agent workflows that require much more working memory.
DeepSeek positions V4-Pro for reasoning, world knowledge, and agentic coding, while V4-Flash is presented as the faster and more economical option. The company also says the models support both thinking and non-thinking modes, while maintaining API compatibility with formats used by OpenAI and Anthropic.
Context
This release matters because AI competition is no longer only about general benchmarks. Long context, inference efficiency, pricing, and the ability to support agents are becoming central criteria for real adoption.
DeepSeek also announced that older `deepseek-chat` and `deepseek-reasoner` models will be retired in July 2026, with current routing moving toward V4-Flash. That suggests V4 Preview is not just a technical demo, but the new base of the company's public offering.
Why it matters
- A 1-million-token context window could reshape workflows involving documents, code, and agents.
- Two variants create a clearer trade-off between capability and cost.
- The launch increases competitive pressure from open-source models on closed frontier labs.
- Immediate API availability makes V4 Preview relevant for practical testing, not just technical discussion.
The important word is `Preview`. DeepSeek is opening its next generation to real-world use, but independent benchmarks, effective costs, and production reliability still need to be tested before treating V4 as a direct replacement for top closed models.