MixtureofMerges-MoE-v2: benchmark results
jsfs11 release with a Mixtral-architecture configuration: 4 local experts, 2 selected per token and 32 decoder layers. The hidden-gate merge recipe uses mlabonne/NeuralBeagle14-7B, jsfs11/TurdusTrixBeagle-DARETIES-7B, FelixChao/WestSeverus-7B-DPO-v2, CultriX/Wernicke-7B-v7. Expert labels do not independently establish total or active learned parameter counts. Provider: jsfs11. Access: Open.
Aggregate Rating 1462 ± 41, rank #3176 of 10916 rated models, from 6 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| Open LLM Leaderboard v1 - ARC Challenge | 72.44 | Normalized accuracy (%) (25-shot) | 92 |
| Open LLM Leaderboard v1 - HellaSwag | 88.41 | Normalized accuracy (%) (10-shot) | 91.6 |
| Open LLM Leaderboard v1 - TruthfulQA MC2 | 70.92 | MC2 (%) (0-shot) | 91 |
| Open LLM Leaderboard v1 - WinoGrande | 83.58 | Accuracy (%) (5-shot) | 89.7 |
| Open LLM Leaderboard v1 - GSM8K | 68.69 | Accuracy (%) (5-shot) | 86.7 |
| Open LLM Leaderboard v1 - MMLU | 64.88 | Accuracy (%) (5-shot) | 77.9 |
Interactive version: theaggregate.ai/model?slug=mixtureofmerges-moe-v2 · How It Works · Data refreshed daily, snapshot 2026-11-08.