Two engine paths. One understandable workflow.

Match the engine to the model

Strata is a specialized engine for supported architectures. Its preparation and expert-offloading workflow can use GPU memory, system memory and storage in ways that differ from an ordinary fully offloaded dense model.

Native llama-server serves compatible GGUF models through llama.cpp. Architecture and accelerator support depend on the actual build you install. A model loading in one engine does not establish compatibility with another.

Keep the boundaries clear

Coding client
    ├─ Model API → selected inference engine → model artifacts
    └─ Tools → approved MCP bridge → local application

The coding client manages conversation and tool calls. The inference engine generates responses. The bridge executes permitted actions in another application.

Budget the whole request

Input, output and tool descriptions share context. Compute a safe output budget before sending a request. Use server tokenization when available, and reserve a buffer for formatting and added messages.

Cached request speed is useful, but it is not the same measurement as uncached prompt processing or total task time. Keep those labels next to performance numbers.