DeepSeek launches open source V4 Preview with 1 million token context

DeepSeek has officially launched DeepSeek V4 Preview, now available as an open source model. The announcement, made via the official X account @deepseek ai, heralds the start of a cost effective era of 1 million token context. The model can be tried directly at chat.deepseek.com in Expert Mode or Instant Mode, and the API has been updated and is available the same day. DeepSeek V4 comes in two main variants. The first is DeepSeek V4 Pro, with a total of 1.6 trillion parameters and 49 billion active parameters. This variant is claimed to deliver performance rivaling the world's best closed source models. The second variant is DeepSeek V4 Flash, with 284 billion total parameters and 13 billion active parameters, positioned as a fast, efficient, and economical option.

Both variants use a Mixture of Experts (MoE) architecture, which activates only a subset of parameters for each request. This approach is key to cost efficiency, as not all parameters need to be computed for every task, unlike dense models that activate all parameters. With 49 billion active parameters out of 1.6 trillion, DeepSeek V4 Pro is designed to reduce computational costs while retaining a large knowledge capacity. The 1 million token context capability is a key feature of DeepSeek V4 Preview. This capacity allows the model to process very long documents in a single session, such as entire books, large codebases, or long conversation transcripts.

DeepSeek calls this the era of cost effective 1 million token context, emphasizing that such large capacity does not have to come with high operational costs. Full technical documentation for DeepSeek V4 Pro has been published as a PDF on Hugging Face, and open weights are available in the official collection on the same platform. This move is consistent with DeepSeek's previous approach of opening access to its models and technical documentation for the research and developer community. The launch comes amid intense competition in the large language model industry, dominated by players such as OpenAI, Google, and Anthropic. DeepSeek's open source strategy offers an alternative for developers who want full control over their models without being tied to proprietary APIs.

The availability of open weights allows developers to fine tune and deploy models on their own infrastructure. From a market perspective, the release of a model claiming performance rivaling the best closed source models could put downward pressure on AI service prices. Open source models with low inference costs are often considered by companies looking to reduce cloud AI spending. The newsletter aigc.news highlighted that the DeepSeek V4 launch was followed by two consecutive price cuts, indicating an aggressive pricing strategy to capture market share. DeepSeek's API documentation confirms the availability of these models for developers through an API update effective today. Developers can access DeepSeek V4 Pro and DeepSeek V4 Flash via the same API endpoints as previous models, with extended context parameters up to 1 million tokens.

This is an added value for applications requiring long document processing, such as legal analysis, academic research, and code analysis. DeepSeek's decision to release the model in Preview status indicates that development is ongoing and a final version may come with further refinements. The preview status also gives DeepSeek room to gather feedback from early users before settling on a stable version. Meanwhile, the availability of two variants of different sizes gives users flexibility to choose between maximum performance and cost efficiency according to their needs. The open source model competition is heating up with the arrival of DeepSeek V4.

A model with 1.6 trillion total parameters places DeepSeek among the very large scale models, while the relatively small number of active parameters shows a focus on computational efficiency. The combination of large scale and low operational cost is a key selling point DeepSeek offers to the market. DeepSeek's move to open source its model and publish the technical report simultaneously demonstrates a commitment to research transparency. This contrasts with the approach of some competitors who keep their model architecture details closed. This openness allows the research community to validate performance claims and explore further development on the foundation provided by DeepSeek. With the API immediately available and the model free to try on the chat platform, DeepSeek is targeting rapid adoption by developers and companies.

The combination of open weights, technical documentation, and direct API access provides three different paths for users to adopt this technology. This strategy is designed to expand the user base while building an ecosystem around the DeepSeek V4 model.