DeepSeek-V4 Arrives with a 1-Million-Token Context and Best-in-Class Open-Source Coding Performance
DeepSeek released previews of its new models, “DeepSeek-V4-Pro” and “DeepSeek-V4-Flash.” Both come with a 1-million-token context length as standard and are available immediately via the API.
Details
- Model lineup: Two models — V4-Pro (1.6 trillion total parameters, 49 billion active) and V4-Flash (284 billion total parameters, 13 billion active)
- Agentic performance: Achieved best-in-class results among open-source models on agentic coding benchmarks
- Strong areas: Performance rivaling top closed-source models in world knowledge, reasoning, math, and STEM fields
- New architecture: Uses token-level compression and DeepSeek Sparse Attention (DSA) to efficiently process long contexts
- V4-Flash: Delivers inference performance close to V4-Pro while responding faster
How to try it
- The API is available immediately, supporting both OpenAI’s ChatCompletions format and Anthropic’s API format
- Also supports integration with existing tools such as Claude Code, OpenClaw, and OpenCode
- The older models (deepseek-chat and deepseek-reasoner) are scheduled to be retired on July 24, 2026, so early migration is recommended