DeepSeek models, providers and cost

DeepSeek first

Start with the official DeepSeek provider because that is what this tutorial evaluates. In Settings -> Models, paste the API key from platform.deepseek.com and select the model for the mission.

Use V4-Flash for most exploration, documentation and controlled-change tasks. Reserve V4-Pro for missions where reasoning failure costs more than the extra price or latency.

• V4-Flash: balance of speed, cost and context.
• V4-Pro: stronger option for difficult tasks, deep reviews or complex plans.
• Thinking/reasoning: raise it only when the task justifies it; measurement matters more than maxing every switch.

How DSH stores credentials

The Models page treats keys as write-only: after saving, the UI receives a redacted descriptor, not the literal secret.

The official guide says the key is stored in `$DSH_HOME/.credentials.yaml`, while settings keep only a credential reference. That changes your audit: review references and file permissions, not screenshots with secrets.

Safe check:
- do not paste API keys into prompts
- do not commit them
- review $DSH_HOME permissions
- rotate the key if it appeared in a screenshot

Additional providers

DSH lets you swap provider without changing the harness. That does not mean you should connect everything: every provider adds a data policy, a bill and a failure surface.

Add OpenAI, Anthropic, Google, Kimi or another provider only when you need to compare a concrete task or cover a clear limitation of the primary model.

Provider decision:
- task DeepSeek does not solve well
- data policy reviewed
- cost per outcome estimated
- same prompt and workspace tested
- decision recorded in the Trajectory

Catalog provider vs custom provider

Add provider uses the installed catalog: endpoint, protocol and model list are predefined for known providers.

Add a custom provider is for a company gateway, self-hosted server or provider absent from the catalog. The Provider ID must be lowercase and is permanent: requests, saved sessions, defaults and credential references use it.

• To rename a provider, add a new provider and delete the old one.
• Display name, base URL, protocol, credential and models remain editable.
• Fetch available models calls compatible `GET /models`; if it fails or does not exist, enter models manually.

Input modalities

A manually entered model is treated as text-only unless declared otherwise, because DSH cannot know which modalities the endpoint accepts.

For vision models on custom providers, the official docs specify `input: [text, image]` in `$DSH_HOME/settings.yaml`. DeepSeek chat-completions is documented as text-only in this guide.

llm-pi-ai:
  providers:
    vision-gateway:
      apiKeyEnv: GATEWAY_API_KEY
      api: openai-completions
      baseURL: https://vision.example/v1
      models:
        - id: vision-preview
          input: [text, image]

Common model errors

Before changing prompts or permissions, diagnose the provider. The official guide names errors worth recognizing.

• `MISSING_CREDENTIAL`: save the key in Models or provide the referenced environment variable.
• `UNKNOWN_MODEL`: select a configured model or add it to the custom provider.
• `GET /models` returns 401: check the credential; if discovery is unsupported, enter models manually.
• Image refused before sending: the model does not declare image input.

Real cost in DSH

Cost is not just token price. In an agent you must count turns, repeated context, cache, failed tool calls, retries and the human review afterwards.

DSH becomes useful when you keep the Trajectory and metrics for every session: input/output tokens, cache hit rate, total time, mode and approval count.

Comparison record:
model:
mode: Standard / Code / Minimal / Creator
exact prompt:
input tokens:
output tokens:
cache hit rate:
total time:
retries:
accepted result: yes / no
estimated cost:

Local models belong to another phase

This phase does not teach installation, network security, GPU checks or local-runtime troubleshooting. Those belong in Aulafy's local AI course.

Inside DSH, the architecture point is narrower: a local model can act as a provider if its server is already working and validated. The server setup is taught separately so the two tutorials stay clean.

• Here: select and measure providers inside DSH.
• Local AI: install, serve models, expose endpoint, GPU and local privacy.
• Phase 8: DeepSeek local-model limits and quantization from the DSH point of view.