DeepSeek V4 – Built for Research, Math, and Long Docs
  • Nano Banana
  • Nano Banana 2
  • Seedream 4.0
  • 30+ AI Models
  • DeepSeek V4: Open-source Intelligence at Scale

    A fundamentally new architecture from DeepSeek. One million tokens of context, world-class performance across coding, math, and agentic tasks, and open weights available to run, fine-tune, or deploy anywhere.

    Trusted by users from 10,000+ companies

    Additional DeepSeek V4 Capabilities

    Architecture innovations and practical capabilities that compound across every type of work.

    Three Modes

    Three Modes

    Switch between Non-Think, Think High, and Think Max modes. Trade speed for depth depending on the task, from fast responses to maximum reasoning.

    Dual Model Tiers

    Dual Model Tiers

    V4-Pro for maximum intelligence and V4-Flash for lightweight, fast, and low-cost use cases. Both share the same 1M context window and reasoning architecture.

    MIT License

    MIT License

    Fully open weights for commercial and research use. Fine-tune, modify, and redistribute without restriction, with no proprietary terms attached.

    Frequently Asked Questions

    Find out the answer to the most common questions people are asking.

    Agent Integrations

    Agent Integrations

    Native integration with Claude Code, OpenClaw, and OpenCode. DeepSeek uses V4-Pro for its own internal agentic coding workflows.

    Dual API Compatibility

    Dual API Compatibility

    Supports both OpenAI ChatCompletions and Anthropic API formats. Drop into any existing integration without changing your client code.

    Context Cache Pricing

    Context Cache Pricing

    Input cache pricing has been cut to one-tenth of previous levels as a permanent change, making repeated long-context calls significantly cheaper at scale.

    Document Generation

    Document Generation

    DeepSeek V4 generates complete, formatted PDFs and slide decks directly from conversational prompts, with layout and structure handled by the model.

    Math and STEM

    Math and STEM

    V4-Pro outperforms all open-source models and rivals the best closed-source models on the most demanding mathematical reasoning evaluations.

    Local Deployment

    Local Deployment

    Fully runnable via vLLM, SGLang, and Docker. Quantized versions available through Ollama and LM Studio for resource-constrained environments.

    Open-source Record, Frontier Competition

    DeepSeek V4 sets a new bar for open-source models across coding, knowledge, and long-context tasks, trading punches with the best closed-source models available.

    Best in Competitive Coding

    Best in Competitive Coding

    DeepSeek V4 scores 93.5% on LiveCodeBench, placing it above every other model in the comparison including Claude Opus 4.6, GPT-5.4, and Gemini 3.1 Pro. It also achieves the highest Codeforces rating of all tested models at 3206.

    Comprehends Long Context

    Comprehends Long Context

    DeepSeek V4 scores 62% on CorpusQA at the full one million token context length. For teams processing full legal filings, research libraries, or large codebases in a single session, it is the strongest open-source option available.

    Strongest Factual Recall

    Strongest Factual Recall

    V4-Pro Max scores 57.9% on SimpleQA-Verified, comfortably ahead of every open-source model and well above closed-source competitors including Claude Opus 4.6 at 46.2% and GPT-5.4 at 45.3%. Only Gemini 3.1 Pro scores higher among all models tested.

    Answers Grounded in Real Knowledge

    DeepSeek V4 leads all open-source models on factual retrieval. For knowledge-intensive work like research and technical analysis users get answers that are grounded in broad, accurate world knowledge rather than confident-sounding generalisations.

    Answers Grounded in Real Knowledge

    Full-document Reasoning and Search

    V4-Pro processes a full one million tokens in a single session and scores ahead of every other model tested. Users working with long legal filings, large research corpora, or full codebase reviews get a model that reasons across the entire document.

    Full-document Reasoning and Search

    Complex Math and Science, Solved Reliably

    For researchers, engineers, and analysts working through quantitative problems, DeepSeek V4 delivers structured, step-by-step reasoning that holds up under scrutiny rather than arriving at plausible-looking but incorrect conclusions.

    Complex Math and Science, Solved Reliably