DeepSeek-V4 Preview Officially Released and Open-Sourced

April 24 Update: DeepSeek has officially launched the preview version of its DeepSeek-V4 model, releasing it simultaneously as an open-source project. Engineered with a massive 1-million-token context window, V4 establishes a new benchmark in the domestic and open-source AI landscape, delivering industry-leading performance in agentic workflows, global knowledge retrieval, and advanced logical reasoning. The model is available in two distinct configurations.

Users can immediately access the DeepSeek-V4 experience via the official web portal at chat.deepseek.com or through the dedicated mobile application to explore its unprecedented 1M context memory. The API infrastructure has been concurrently upgraded; developers can seamlessly integrate the latest models by updating the model_name parameter to either deepseek-v4-pro or deepseek-v4-flash.

Key specifications for each variant are outlined below:

• Enhanced Agentic Capabilities: The DeepSeek-V4-Pro delivers a substantial leap in autonomous agent performance. In Agentic Coding benchmarks, it achieves state-of-the-art results among open-source models and excels across related evaluation metrics. Already deployed internally for enterprise coding workflows, user feedback indicates a superior experience compared to Sonnet 4.5, with output quality closely matching Opus 4.6 in standard mode, though a measurable gap remains against its advanced reasoning variant.

• Comprehensive World Knowledge: In extensive knowledge assessments, DeepSeek-V4-Pro significantly outpaces competing open-source alternatives, ranking just behind top-tier proprietary systems like Gemini-Pro-3.1.

• World-Class Reasoning Performance: Evaluations spanning advanced mathematics, STEM disciplines, and competitive programming demonstrate that V4-Pro surpasses all publicly benchmarked open-source models, achieving performance parity with elite closed-source AI systems.

DeepSeek-V4-Flash

• Optimized for efficiency, the Flash variant features a slightly reduced knowledge footprint while maintaining reasoning capabilities nearly identical to the Pro edition. Its streamlined parameter count and lower activation overhead translate to significantly faster response times and more cost-effective API consumption.

• In standardized agent evaluations, V4-Flash performs on par with V4-Pro for routine tasks, though it exhibits a performance differential when handling highly complex, multi-step operations.

Under the hood, DeepSeek-V4 introduces a novel attention architecture that compresses data at the token level. Coupled with DeepSeek Sparse Attention (DSA), this breakthrough delivers unparalleled long-context processing while drastically reducing computational and VRAM requirements compared to conventional approaches. Moving forward, a 1M context window will serve as the standard baseline across all official DeepSeek services.

The model has been rigorously optimized for seamless integration with leading agentic development frameworks, including Claude Code, OpenClaw, OpenCode, and CodeBuddy. These enhancements yield marked improvements in code generation, technical documentation, and automated workflow execution. (Official benchmarks include a sample slide deck auto-generated by V4-Pro within a specific agent environment.)

Both V4-Pro and V4-Flash support a maximum context length of 1M tokens and offer dual operational modes: standard and deep reasoning. The reasoning mode features a configurable reasoning_effort parameter (high / max). For intricate agentic scenarios, activating deep reasoning with the intensity set to max is highly recommended.

Additionally, legacy API identifiers deepseek-chat and deepseek-reasoner will be officially deprecated on July 24, 2026. During the transition period, these endpoints will automatically route to the standard and deep reasoning modes of deepseek-v4-flash, respectively.