> ## Documentation Index
> Fetch the complete documentation index at: https://documentation.deepmask.io/llms.txt
> Use this file to discover all available pages before exploring further.

# DeepSeek

> DeepSeek V3 and V3.1 (Infercom) on DeepMask. 671B MoE architecture delivering frontier-class coding, math, and document analysis without image support.

DeepSeek's V3 family redefines the cost-to-intelligence ratio using a 671B parameter Mixture-of-Experts architecture. DeepSeek V3 sets a new standard for efficient coding and STEM reasoning, while DeepSeek V3.1 (Infercom) builds on that foundation with a hybrid thinking mode and EU-hosted infrastructure via Infercom. Neither model supports image input, making both ideal for text and document-heavy workflows.

<Tabs>
  <Tab title="DeepSeek V3">
    ## About

    DeepSeek V3 is a 671B parameter Mixture-of-Experts model that has set a new industry standard for efficiency. Its innovative Multi-head Latent Attention (MLA) architecture delivers frontier-class coding and math performance at a fraction of the hardware cost. It is widely regarded as the best model for developers who need maximum reasoning power at the lowest possible price.

    <Note>
      DeepSeek V3 does not support image inputs. Use it for text, code, and document-based tasks.
    </Note>

    ## Key Capabilities

    <CardGroup cols={2}>
      <Card title="Mathematical Proofs" icon="square-root-variable">
        Outperforms most frontier models on the AIME and MATH-500 benchmarks for complex symbolic reasoning.
      </Card>

      <Card title="Cybersecurity Awareness" icon="shield">
        Highly effective at identifying vulnerabilities in C++, Rust, and Python codebases.
      </Card>

      <Card title="Consistent Logic" icon="check-double">
        Delivers highly consistent reasoning across all query types, reducing unexpected output drift.
      </Card>

      <Card title="Fast Response" icon="bolt">
        Efficient decoding architecture accelerates response times without losing precision.
      </Card>
    </CardGroup>

    ## Use Cases

    * **Low-cost coding agents** — Build production-grade code generators and automation pipelines with minimal per-task cost.
    * **STEM research** — Solve complex engineering problems and symbolic math equations at scale.
    * **Bulk data transformation** — Reformat and clean massive datasets with structural precision.
    * **Document analysis** — Extract structured information from dense technical or legal documents.

    <Tip>
      DeepSeek V3 is your best choice when you need strong logic and coding ability at the lowest token cost. For EU-hosted deployments or hybrid thinking modes, use DeepSeek V3.1 (Infercom).
    </Tip>

    ## Specifications

    | Specification    | Value                                         |
    | ---------------- | --------------------------------------------- |
    | Model Provider   | DeepSeek                                      |
    | Main Use Cases   | High-Efficiency Agents, STEM, Bilingual Logic |
    | Reasoning Effort | Adaptive (Non-Thinking / Thinking)            |
    | GPQA Diamond     | 80.7%                                         |
    | Max Context      | 128K – 164K Tokens                            |
    | Latency (TTFT)   | 0.41s                                         |
    | Throughput       | 74 Tokens/sec                                 |
  </Tab>

  <Tab title="DeepSeek V3.1 (Infercom)">
    ## About

    DeepSeek V3.1 (Infercom) is the updated "Terminus" release of the DeepSeek V3 family, refined for high-scale managed APIs (MaaS). It is a hybrid model that supports both a high-speed "Non-Thinking" mode for general chat and a deep "Thinking" mode for complex reasoning. The Infercom variant is specifically optimized for sub-second responses in autonomous agent and API integration scenarios.

    <Note>
      DeepSeek V3.1 is EU-hosted via Infercom, making it a strong option for organizations with European data residency requirements. Image input is not supported.
    </Note>

    ## Key Capabilities

    <CardGroup cols={2}>
      <Card title="Dual-Mode Inference" icon="toggle-on">
        Toggle between fast non-thinking mode for quick answers and deep thinking mode for complex multi-step logic — all in one model.
      </Card>

      <Card title="Faster Reasoning" icon="bolt">
        The 3.1 update reduced time-to-answer for reasoning queries by 30% compared to earlier iterations.
      </Card>

      <Card title="Math & STEM Dominance" icon="square-root-variable">
        Achieves 93.1% on AIME 2024, making it a price-performance leader for technical problem-solving.
      </Card>

      <Card title="Structured Data Extraction" icon="table">
        Reliable document-to-JSON conversion with high accuracy for data pipeline workflows.
      </Card>
    </CardGroup>

    ## Use Cases

    * **High-volume API integration** — Provide smart reasoning for thousands of simultaneous users at a fraction of the cost of US-based models.
    * **Bilingual RAG** — Excel at English-Chinese technical documentation and cross-border business intelligence.
    * **Structured data extraction** — Run reliable document-to-JSON pipelines with the Infercom managed API.
    * **Autonomous agents** — Deploy cost-efficient agentic loops that alternate between fast chat and deep reasoning as needed.

    <Tip>
      Choose DeepSeek V3.1 (Infercom) when you need EU data residency, hybrid thinking modes, or high-throughput API deployments. Its 32K tokens/sec throughput makes it one of the fastest options available on DeepMask.
    </Tip>

    ## Specifications

    | Specification    | Value                                 |
    | ---------------- | ------------------------------------- |
    | Model Provider   | DeepSeek                              |
    | Main Use Cases   | Bilingual API Dev, Low-Cost Reasoning |
    | Hosting          | EU-hosted via Infercom                |
    | Reasoning Effort | Hybrid (Think / Non-Think)            |
    | GPQA Diamond     | 93.1%                                 |
    | Max Context      | 164K Tokens                           |
    | Latency (TTFT)   | 0.21s                                 |
    | Throughput       | 32K Tokens/sec                        |
  </Tab>
</Tabs>