Models & Providers

Pioneer supports local Ollama models and configured cloud or OpenAI-compatible providers. The product keeps durable intent separate from live provider evidence.

Three model states

State	Meaning
Configured	The provider and default model saved in the active profile.
Effective	The configured choice currently resolves to a usable Chat model under Pioneer's provider checks.
Observed	Inventory, capabilities, reachability, residency, or context reported by the provider and Machine evidence.

A removed local model can remain configured while no effective model exists. Pioneer does not invent a fallback and does not present the configured name as observed inventory.

Where model settings belong

• Settings owns provider endpoints, credentials, preferences, and advanced provider configuration.
• Models owns the durable default, provider tests, local model lifecycle, installed inventory, and context evidence.
• Machine owns observed runtime reachability, residency, and hardware state.
• The Chat model picker selects a model for the conversation. It does not rewrite the profile default.

Cloud credentials are write-only. Pioneer can reuse a stored credential for the same provider and endpoint without returning the secret to the browser.

Ollama model lifecycle

In Models, you can explicitly install, prewarm, unload, or delete supported local models.

• Install is a network and disk mutation. It does not set the default or load the model.
• Set as default is a separate durable configuration write.
• Prewarm 5 min requests temporary residency without sending a user prompt.
• Unload now requests immediate unload; it does not prove that active inference was stopped because Pioneer cannot authoritatively observe that.
• Delete removes model files after destructive confirmation. It is blocked while the model is the configured/effective default or observed as resident.

Pioneer never installs a model merely because it appears in configuration, onboarding, or a Chat request.

An installed model can still be excluded from Chat when provider evidence says it is embedding-only or when completion capability could not be resolved. The Chat and onboarding pickers include only eligible completion models; Models keeps the broader inventory visible so the exclusion can be diagnosed.

Installed, resident, and active

• Installed means the model appears in Ollama inventory.
• Resident means Ollama reports a loaded model process.
• Active generation is not exposed authoritatively by the current Machine contract and must not be inferred from residency.

Context evidence

Pioneer distinguishes:

• model capacity reported by provider metadata;
• runtime-loaded context observed from a resident process;
• runtime context requested for generation; and
• the usable Chat session budget after practical caps and reserves.

If capability probing fails, capacity remains unavailable. Pioneer keeps a prior safe session allocation rather than fabricating a model capacity.

Provider checks

The provider test is bounded and non-generative. It checks reachability and model discovery without sending a Chat prompt. A real First Chat is a separate, optional operation that creates durable session and transcript evidence.

For current symptoms and checks, see Troubleshooting.

Choosing a model

Choose from current observed capability, required context, privacy constraints, latency, and the work you are doing. A benchmark result is evidence about one tested setup, not a permanent universal default. Pioneer remains usable across models and does not make the manual depend on a single local model name.

Changing the durable default affects later default selection. Changing the Chat composer model targets that conversation after capability preflight. Configuring a model_call block records intent for that block; see the Model Call reference for its execution evidence limitation.