The previous three episodes covered how to evaluate open-weights models, how to access them through providers, and what it takes to run them on your own hardware.
The last piece is the harness — the software that sits between you and the model.
A harness can be very simple or quite sophisticated. In practice people move through three main layers.
The first is a minimal chat interface such as llama.cpp. You point it at a model file and start a chat. There is no conversation history beyond the current session, no tool calling, and no isolation from the rest of the system. It is useful for a quick test, especially after downloading a quantized model, because setup is minimal and feedback is immediate.
The second layer is the more complete command-line interfaces. Examples include the CLIs from Anthropic, OpenAI’s Codex-style CLI, and Grok’s own build tooling. These add three practical features: sandboxing (the model does not operate directly on your files), tool use (the model can call external functions such as web search or file searching), and session management (you can pause and resume work, sometimes even from a phone). For people who already live in the terminal (guilty as charged), this is often the daily driver.
The third layer consists of much more complex interfaces such as Pi or Hermes. These provide graphical or conversational UIs, better support for spoken interaction, and in some cases even connections to messaging apps. They feel closer to a personal assistant. Non-technical users often prefer this level once they decide they want something more polished.
Most people do not need to start at the top. A basic tool is enough to decide whether a particular model is worth keeping. Once daily use begins, a proper CLI usually becomes the practical choice. The more elaborate interfaces are optional and mainly about comfort and reach.
That completes the short series. The four episodes together give an easy way to evaluate: the metrics that describe a model, the two ways of accessing it, the hardware choice for local inference, and the software layer that makes the model usable.
The full playlist is here: https://youtube.com/playlist?list=PLKxL2crAoLBU&si=3nDQeNgCw51rFDuG
Top comments (0)