Launching DSH from Ollama
Recent Ollama releases add a launch command for DeepSeek Harness, so the harness can be started against a locally served model without a hosted API key. Check your installed Ollama version first, since the command is only present in newer builds.
ollama --version
ollama launch dshPick the model before the plugins
The model you serve locally determines the capability floor for everything above it. Tool-calling reliability and context length matter more than raw parameter count for agent work.
- Confirm the model supports tool or function calling.
- Check the context length against your typical session size.
- Expect slower iteration than a hosted frontier model.
- Verify quantization has not degraded structured output.
What changes for plugins on a local model
Plugins that only add tools, interfaces, or storage behave the same. Plugins that assume a specific model capability do not.
| Plugin type | On a local model | Why |
|---|---|---|
| Terminal and Web interfaces | Unaffected | They sit above the model adapter |
| Browser and sandbox tools | Usually fine | They depend on tool calling, not modality |
| Memory and retrieval | Usually fine | Storage is model-independent |
| Vision bridges | Depends | A text-only local model can reject attachments before a session sees them |
| Model-specific adapters | Often broken | They assume a provider-specific API surface |
Keep the local setup reproducible
A local stack has more moving parts than a hosted one, so record the combination that worked.
- Note the Ollama version, model tag, and quantization.
- Keep local-model work in its own DSH profile.
- Pin plugin versions so a model change is the only variable.
- Re-test tool registration after every model swap.
DeepSeek Harness is in developer preview. Recheck official documentation and plugin evidence when the host version changes.
Official sources
These primary sources were used to verify the technical details in this guide.