Local-model inference runs a trained model on hardware controlled by the user or deploying organization. Reload a base model with an adapter, optionally merge and export it, then serve requests without sending sensitive data to a remote service. This preserves confidentiality, avoids network latency, and keeps control over model outputs. section 2.2
A practical local deployment workflow is:
- Save adapter weights and reload the base model with the selected adapter. section 2.2
- Create an Ollama Modelfile, register the model locally, and run it offline; use llama.cpp directly for a more transparent toolchain. section 2.2
- Use LM Studio when graphical experimentation or CPU/GPU execution is preferred. section 2.2
Offline financial assistant
Declare a finance adapter with FROM llama2 and ADAPTER ./finance_adapter, then run ollama create finance-model -f Modelfile followed by ollama run finance-model. The resulting assistant performs financial analysis locally without external calls. section 2.2

