Compact model experiments focused on capability per parameter and visible API cost.
Capability per stage, not per product
A research product does not need one model to perform every operation. Query classification, extraction, deduplication, citation formatting, and final synthesis have different capability requirements.
North experiments with sub-billion-parameter models for bounded stages where latency, memory, and repeatability matter more than broad world knowledge. The parameter count is a constraint, not a quality claim.
Cost accounting that users can see
Total API cost is shaped by input length, retrieval volume, number of passes, model choice, and retries. A low token price can still produce an expensive workflow if the system repeatedly resends the same context.
Horizon budgets work by stage and preserves intermediate artifacts so a failed synthesis does not force retrieval and extraction to run again. The target is predictable research economics, not a universal promise of a particular price.
Where compact models fail
Small models are less forgiving when instructions are ambiguous, domains shift, or the task requires long-range synthesis. Routing must detect those cases and escalate rather than forcing a compact model to imitate capability it does not have.
Evaluation therefore includes refusal quality, escalation accuracy, and sensitivity to prompt variation. Cheap failure is still failure; efficiency only matters when the result remains useful.