DeepSeek

Last updated: 2026-10-01. This page describes the bundled implementation, not a remotely queried model catalog. Check your account's available models before deployment.

Implementation

The deepseek plugin uses the shared Chat Completions adapter. Its dialect supplies https://api.deepseek.com with /chat/completions, bearer authentication, thinking/reasoning handling, reasoning-content replay, and cache-hit usage normalization. Provider execution and retries remain inside the model layer.

The plugin advertises these local runtime options:

Name	Choices
model	deepseek-flash, deepseek-v4-pro
reasoning_effort	low, high, max

These are the choices returned by get_options(), not a statement that every account or compatible third-party endpoint accepts them. Startup model and provider-specific config are distinct from the validated runtime option patch. The endpoint remains configurable for a compatible service.

Worker configuration

Copy this provider section into your worker YAML. Shared endpoint, authentication, and retry semantics are described in Support policy.

providers:
  deepseek:
    plugin: deepseek
    endpoint:
      base_url: https://api.deepseek.com
      request_path: /chat/completions
      auth:
        scheme: bearer
        api_key: ${DEEPSEEK_API_KEY}
        header_name: x-api-key  # Used only with scheme: custom_header.
        extras: {}             # Metadata; not sent by this plugin.
      user_agent: simplex-cpp/deepseek
      extra_headers: {}
      extras: {}               # Metadata; not generation parameters.
    model: deepseek-flash      # Also advertised: deepseek-v4-pro.
    config:
      reasoning_effort: high   # Public runtime choices: low, high, max.
      thinking:
        type: enabled          # Native toggle: enabled or disabled.
      # Alternative to reasoning_effort, not an additional control:
      # reasoning:
      #   effort: high
      # Optional passthrough examples; values are not plugin defaults:
      # max_tokens: 8192
      # temperature: 1.0
      # top_p: 1.0
      # stop: ["END"]
      # response_format: {type: json_object}
      # tool_choice: auto
      # logprobs: true
      # top_logprobs: 5
    retry:
      max_attempts: 3
      initial_backoff_ms: 500
      max_backoff_ms: 120000

The endpoint, authentication scheme, and user agent above are the plugin's transport defaults. The model and explicit generation settings are example choices. If both thinking controls are omitted, the plugin enables thinking but does not supply a reasoning effort.

Control	Plugin behavior
thinking	A native object is forwarded unchanged and takes precedence over the local effort-based thinking toggle. Use type: disabled and omit the effort to disable thinking explicitly.
reasoning_effort	low, high, and max are advertised to clients. Startup configuration is permissive; other values reach the service unless consumed by the local toggle.
reasoning.effort	Shared compatibility spelling. Converted to reasoning_effort; an explicit top-level value wins. The reasoning envelope is removed.
none / minimal effort	Legacy local disable aliases: without a native thinking object, disable thinking and remove the effort from the request. They do not override an explicit native object.
n, frequency_penalty, presence_penalty	Removed before transmission; do not use them as effective settings.

Keep native thinking and effort settings consistent. Sampling, output limits, structured output, and other passthrough parameters are validated by the remote service; some depend on thinking mode. For JSON output, also instruct the model to produce JSON. Consult the DeepSeek request reference for current server-side constraints.

Select this configuration with driver_model: deepseek; optionally set modality_assist_model: deepseek for a separate image-assistance instance. Export DEEPSEEK_API_KEY in the environment that actually starts the worker. The loader expands credentials only for selected model configurations. For a full worker file, see the minimal configuration, which names its provider entry primary instead.

Use from the Hub

Save the worker YAML in the configuration library. Select it together with a local or Docker launch file when creating a session. The Hub fills worker connection endpoints but preserves provider configuration. For an existing session, stop it, apply the saved configuration pair, and restart it.

The panel obtains advertised model choices through the worker's options signal. Selected changes are attached to the next run-starting payload; they do not mutate an active run. Runtime selections are not persisted as startup configuration in AgentInputState.

Optional image assistance

modality_assist_model: deepseek

This creates a separate model instance even when both roles select deepseek. It enables an intrinsic tool that reads local image files, encodes them as base64, and performs one isolated model exchange to return a textual description. The endpoint/model must support the resulting image input. A driver invocation can call this tool; merely setting the role does not automatically convert every incoming attachment. Every supplied image path must succeed or the call fails.

Failure diagnosis

Check worker logs for missing credentials, incompatible plugin libraries, endpoint/authentication failures, and unsupported provider requests. Exhausted model retries are reported as failed runs; continuation is possible only when the settled state permits it. Provider errors must not be mistaken for lost conversation state or proof that earlier tool effects were undone.