FOSCAR - Qwen3-8B - GPQA Diamond: leaderboard
FOSCAR v1 Table2 accuracy (%), Qwen3-8B under three native KV-cache configurations FlashInfer BF16, OSCAR INT2 and OSCAR INT2+FOSCAR. These are inference configurations, not distinct released checkpoints. Retain printed mean; uncertainty remains in pinned source table. No throughput, latency or kernel resource metric is admitted.
Metric: Accuracy / task score (%). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 3 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen3-8B (FOSCAR Run: cache=FlashInfer BF16;effort=source-unreported) | 56.36 | |
| 2 | Qwen3-8B (FOSCAR Run: cache=OSCAR INT2 + Ours;effort=source-unreported) | 55.45 | |
| 3 | Qwen3-8B (FOSCAR Run: cache=OSCAR INT2;effort=source-unreported) | 54.65 |
Interactive version: theaggregate.ai/benchmark?slug=foscar-qwen3-8b-gpqa-diamond · How It Works · Data refreshed daily, snapshot 2026-11-12.