Skip to content
Bedrova bedrova
FeaturesCompatibilityDevelopersCompareDocs Coming soon

Docs

Model profiles & presets

Per-model settings layered over your global defaults, effective-settings previews, and the portable bedrova.model.json sidecar.

Every model carries a profile — its own settings, layered over your global defaults — so a reasoning model and a fast 7B can each behave the way you want without you re-tuning every chat.

What a profile holds

A model’s profile can set:

  • System prompt — the default instruction for this model.
  • Load settings — context length and load-time options.
  • Sampling — temperature, top-p, top-k, min-p, penalties.
  • Reasoning — effort mode and whether reasoning is shown.
  • Structured output — a default response schema (see Structured output).
  • Speculative decoding — draft model and whether it’s on (see Performance).
  • Prompt template — the chat template, when a model needs a specific one.
  • RAG defaults — retrieval behavior for document chat.

Effective settings

What actually applies to a request is the effective settings — your global defaults, with the model’s profile layered on top, with any per-chat override on top of that. The app previews the effective values as you change them so there’s no guessing about precedence:

global default → model profile → this chat
                                     ↳ effective settings

Set global defaults in Settings → Model Defaults; override per model in its profile; override once in a single chat without changing either.

Reset

Any profile resets to your global defaults — per setting or the whole profile — so experimenting is safe. A reset profile simply falls back to the defaults again.

The portable sidecar

A model’s metadata, preferences, profile, benchmarks, local overrides, Hugging Face provenance, and compatibility notes can travel with it as a portable bedrova.model.json file. This is how a profile follows a model:

  • Import reuses an existing bedrova.model.json when you add a local model folder, so your tuning comes along.
  • Export writes the sidecar so you can move a tuned model to another Mac or back it up.

Local model import and the sidecar are covered in Import local models; the schema is the shared portable-model-metadata.v1 contract.

See also