Inkling 2 Small

CymraegBench v0.1 — low effort open weights

2,053 cases completed: no invalid outputs, refusals, infrastructure errors or token-limit truncations.

Protocol and scoring

Mandatory low-effort reasoning; corrected binary-answer extractor (binary-boundary-v2). This is not a thinking-off run.

OpenRouter, Nebius provider; temperature 0 and seed 1. A 2,048-token completion floor shared between reasoning and the answer. Raw responses are retained. Explicit Y/N labels at the start or on the final line are accepted; conflicting labels are rejected. Not all older results have been rescored with this extractor.

Despite the name “Small”, this is a 594B-parameter MoE with 17B active parameters. It belongs in the main open-model table, not the up-to-7B group. Model card.

CCC Language Audit

93.37 — mean of the seven subsets, separate from the overall score.

Data

• leaderboard.csv
• suite-scores.csv
• ccc.csv
• metadata.json

Back to results