Ornith-1.5-397B: benchmark results
Ornith AI's open-weight Ornith-1.5-397B (August 2026), the flagship 397B MoE model in the Ornith-1.5 family, extending Ornith-1.0's Qwen3.5/Gemma-4 base with a self-improving RL loop. Provider: Ornith. Released 2026-08-19. Access: Open.
Aggregate Rating 1710 ± 23, rank #331 of 2170 rated models, from 34 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| Humanity's Last Exam (Self-Reported, No Tools) | 44.6 | Accuracy (%) | 98.2 |
| Nejumi 4 - ALT - Truthfulness | 85.65 | Score (%) | 94.1 |
| Nejumi 4 - GLP - Semantic Analysis | 85.6 | Score (%) | 93 |
| Nejumi 4 - ALT - Ethics & Morality | 97 | Score (%) | 89 |
| Humanity's Last Exam (Self-Reported, With Tools) | 56.1 | Accuracy (%) | 87.9 |
| Nejumi 4 - ALT Average | 88.69 | Score (%) | 84.6 |
| Nejumi 4 - MT-Bench (Japanese) | 97.19 | Score (%) | 83.8 |
| Nejumi 4 - GLP - Basic Language | 85.22 | Score (%) | 80.1 |
| Nejumi 4 - Total Score | 81.61 | Score (%) | 80.1 |
| BenchLM | 62.5 | Overall Score | 79.5 |
| Nejumi 4 - GLP - Application Development | 65.17 | Score (%) | 78.7 |
| Nejumi 4 - ALT - Bias | 91.22 | Score (%) | 78.3 |
Interactive version: theaggregate.ai/model?slug=ornith-1-5-397b · How It Works · Data refreshed daily, snapshot 2026-11-05.