Reasoning modes
Some Starfire models can spend additional compute on internal reasoning before producing the visible response. Reasoning configuration changes the tradeoff between speed, depth, and usage.
When deeper reasoning helps
Use higher reasoning effort for work such as:
- complex debugging
- multi-step architecture
- difficult math or logic
- constraint-heavy planning
- codebase-wide changes
- research synthesis
Routine rewriting, simple extraction, or short factual questions often do not need the deepest mode.
Reasoning is not verification
A model can reason carefully and still begin from incomplete or incorrect information. Use search, Knowledge, validators, or source inspection when the answer must be grounded in external evidence.
Latency and credits
More reasoning can increase response latency and compute consumption. The exact credit effect is controlled by the current model and metering configuration.
FORGE, Research, and agent workflows can combine reasoning with tools. In those cases, model reasoning is one part of the process; validation, retrieval, or external sources can be more important than simply increasing reasoning effort.
Availability
Reasoning controls only appear for models and accounts that support them. A route without reasoning support cannot gain the capability because a prompt asks for it.
Increase reasoning when the problem is hard, not merely because the output should be long.
