DeepSeek-V4 Preview Release

DeepSeek has released the DeepSeek-V4 Preview, an open-source model series featuring a standard 1M token context length and high-efficiency architecture. This release introduces two distinct model variants: DeepSeek-V4-Pro and DeepSeek-V4-Flash, designed to balance high-end reasoning and cost-effective performance.

Model Variants and Specifications

DeepSeek-V4 introduces two models tailored for different performance and efficiency needs:

  • DeepSeek-V4-Pro: A high-capacity model with 1.6T total parameters and 49B active parameters. It is designed to rival top closed-source models in performance.
  • DeepSeek-V4-Flash: A smaller, more efficient model with 284B total parameters and 13B active parameters, optimized for speed and cost-effectiveness.

Technical Capabilities and Benchmarks

DeepSeek-V4-Pro is positioned as a state-of-the-art open-source model across several key domains:

  • Agentic Coding: The model achieves state-of-the-art (SOTA) performance in agentic coding benchmarks.
  • World Knowledge: DeepSeek-V4-Pro leads all current open models in world knowledge, trailing only Gemini-3.1-Pro.
  • Reasoning: The model outperforms all current open models in Math, STEM, and Coding, rivaling top closed-source alternatives.

DeepSeek-V4-Flash provides reasoning capabilities that closely approach those of the Pro version, and it performs on par with V4-Pro on simple agent tasks, while offering faster response times and lower API pricing.

Structural Innovation and Context Efficiency

DeepSeek-V4 implements novel attention mechanisms to achieve a 1M token context window as the default across all official services. This efficiency is achieved through:

  • Token-wise compression
  • DSA (DeepSeek Sparse Attention)

These innovations reduce the compute and memory costs associated with processing long-context windows, establishing a 1M context standard for the series.

Agent Integration and Optimizations

DeepSeek-V4 is optimized for agentic workflows and is already integrated with leading AI agents such as Claude Code, OpenClaw, and OpenCode. The model is also used internally at DeepSeek for agentic coding tasks.

API Availability and Migration

DeepSeek-V4 models are available via API today, supporting OpenAI ChatCompletions and Anthropic APIs. Users can migrate by updating the model name to deepseek-v4-pro or deepseek-v4-flash while maintaining the same base_url.

Both models support dual modes: Thinking and Non-Thinking.

Important Migration Note: The deepseek-chat and deepseek-reasoner models will be fully retired and inaccessible after July 24, 2026, at 15:59 UTC. Currently, these legacy models are routing to deepseek-v4-flash in non-thinking and thinking modes, respectively.

Sources

Related

  • Dispatch
    DeepSeek-V4 Release Notes: Efficient 1M-Token Context for AI Agents

    DeepSeek-V4 introduces a 1M-token context window powered by a hybrid CSA/HCA attention mechanism, specifically optimized for long-running agentic workloads and tool-use trajectories.

  • Dispatch
    DeepSeek-V4-Flash Update

    DeepSeek has released a public beta update for DeepSeek-V4-Flash, significantly enhancing agent capabilities and benchmark performance through re-post-training while maintaining the original model architecture.

  • Dispatch
    DeepSeek-V4-Pro GA Release

    DeepSeek has released DeepSeek-V4-Pro, featuring enhanced agent capabilities, flexible reasoning effort settings, and native OpenAI Responses API support.

  • Dispatch
    DeepSeek-V4-Flash-Vision-Exp Release Notes

    DeepSeek has released DeepSeek-V4-Flash-Vision-Exp, an experimental multimodal model that bridges the gap between fast text processing and high-performance visual understanding for agentic workflows.