The company reports that Jev and pplx-decider-v1 were the most consistent across tasks, while open models matched Jev on some zero-shot tasks but introduced self-hosting complexity. On the held-out reranking test, Parallel’s fine-tuned Qwen3 0.6B reranker reached 75.2 top-K overlap, versus 71.5 for Jev. For page-change classification, pplx-decider-v1 scored 93.3 AUPRC before prompt tuning; five rounds of prompt tuning lifted Rune 26B from 74.1 to 83.9 on the held-out set. These are results from Parallel’s own production-task evaluation, not a universal ranking. The authors recommend testing models against a labeled sample of the intended task.
Are all System One models created equal? | Parallel
In one minute
Parallel published a benchmark of eight System One decision models—three hosted APIs and five open-weight models—on four production tasks: search-result reranking, page-change classification, low-quality-content filtering, and query routing.