CognitiveFusion-4x7B-bf16-MoE: benchmark results
Kquant03's mixture-of-experts release whose published Mixtral configuration specifies 4 local experts, 2 selected per token and 32 decoder layers. Its recipe names SanjiWatsuki/Loyal-Macaroni-Maid-7B as the base; nominal expert sizes are not measured total or active parameter counts. The card credits collaboration with NeuralNovel. Provider: Kquant03. Access: Open.
Aggregate Rating 1434 ± 34, rank #4493 of 10916 rated models, from 6 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| Open LLM Leaderboard v1 - TruthfulQA MC2 | 67.05 | MC2 (%) (0-shot) | 85.4 |
| Open LLM Leaderboard v1 - MMLU | 65.14 | Accuracy (%) (5-shot) | 82.3 |
| Open LLM Leaderboard v1 - HellaSwag | 86.16 | Normalized accuracy (%) (10-shot) | 79.1 |
| Open LLM Leaderboard v1 - ARC Challenge | 67.41 | Normalized accuracy (%) (25-shot) | 76.7 |
| Open LLM Leaderboard v1 - WinoGrande | 78.69 | Accuracy (%) (5-shot) | 64.9 |
| Open LLM Leaderboard v1 - GSM8K | 52.54 | Accuracy (%) (5-shot) | 64.4 |
Interactive version: theaggregate.ai/model?slug=cognitivefusion-4x7b-bf16-moe · How It Works · Data refreshed daily, snapshot 2026-11-08.