Choose the AI model that reads your training code

Choose which AI model Sigmora uses: a cloud provider through an API endpoint, a model running on your own PC, or the built-in model. Press Test to check an endpoint before you start working on your training scripts.

Choose where your model runs

Cloud AI provider

A cloud AI provider is the choice when a step needs a bigger model, for example when you debug a long training script that spans many files. You enter a standard API endpoint and your API key. Keys stay on your PC. Your prompts travel over the network to that provider, and so does the text of any file the agent reads, so use it only with code and data you are allowed to send there. Secrets in tool output are redacted before they reach the model. Privacy depends on the provider you pick.

Local PC model

A model running on your own PC is the choice when your code or data must not leave the machine. You point Sigmora at a local endpoint on the same computer, and your prompts go there instead of to a provider. With a local model nothing leaves the machine. Speed and quality depend on your hardware and on the model you load, so try it on a small script first. A weak model can miss details that a bigger one would catch, and you review every diff either way.

Built-in model

The built-in model is ready to use without your own endpoint. Install Sigmora, open a chat and give it a task. It may suit a first try, such as writing a small perceptron file, adding a log line to a training script or explaining an error message. For a harder step, such as a change that touches several files, you can connect a bigger model or send that step to one with run_on_model. You can change models at any time, and you read the diff before you accept it.

Test your connection

The Test button checks an endpoint before you rely on it. Press it after you enter the address of a cloud provider or of a model on your PC. Sigmora shows the response time and the context window of that endpoint. The context window tells you how much text, such as a training script and a log together, the model can read in one request. Test each endpoint once. If you have 2 endpoints, press Test on both and compare them. A slow response is worth knowing about before a long chat, and a small context window means you should point the agent at one file at a time. Test checks the endpoint only; it does not rate the quality of the answers.

Pick the right model size

A bigger model is the better choice for hard steps, such as untangling a training loop that spreads across many files or finding out why a loss does not go down. A smaller model may be enough for small edits, such as renaming a variable or adding a log line. You can mix them: run the chat on a smaller model and use run_on_model to send one hard step to a bigger one. Any model can be wrong. Start small. Switch to a bigger model when a step keeps failing, and review the diff each time before you accept it.

Quality varies between models and between providers, so treat every answer as a draft. In Ask first mode Sigmora pauses before actions that change something and shows the diff, so you can read the exact lines before you accept them. Then run your test command with Shell and look at the exit code and the printed numbers yourself. Auto-run acts at once, which suits small steps you already trust. Review first.