DeepSeek V4 Preview goes live with open weights, 1M context, and two model variants

DeepSeek says DeepSeek-V4 Preview is now live and open-sourced, with two model variants: DeepSeek-V4-Pro and DeepSeek-V4-Flash. The company says both models support a 1 million token context window and are available through updated API access starting immediately.

According to DeepSeek, V4-Pro uses a 1.6T total-parameter architecture with 49B active parameters, while V4-Flash uses 284B total parameters with 13B active parameters. DeepSeek says the Pro model is aimed at top-tier reasoning, world knowledge, and agentic coding performance, while Flash is positioned as the faster and more economical option.

The company also says the release introduces a new default 1M-context standard across official DeepSeek services and includes support for both Thinking and Non-Thinking modes. DeepSeek notes that existing `deepseek-chat` and `deepseek-reasoner` endpoints will be retired after July 24, 2026, with traffic currently routed to V4-Flash modes in the meantime.

DeepSeek further says V4 is already integrated with agent frameworks including Claude Code, OpenClaw, and OpenCode, and that it is being used internally for agentic coding work. The release is both a model update and an API migration notice for existing DeepSeek users.