luxia-21.4B-alignment-v1.0: benchmark results
Korean firm Saltlux's 21.4B SFT+DPO alignment model built on a Luxia base depth-upscaled from internlm2-20b with continual pretraining (March 2024). Provider: Saltlux. Released 2024-03-12.
Aggregate Rating 1456 ± 18, rank #965 of 1656 rated models, from 24 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| Open Chinese LLM - HellaSwag | 80.26 | Accuracy (%) | 100 |
| Open Chinese LLM - TruthfulQA MC | 70.67 | Accuracy (%) | 100 |
| Open Chinese LLM - WinoGrande | 74.35 | Accuracy (%) | 100 |
| Open Chinese LLM - ARC Challenge | 69.97 | Accuracy (%) | 99.1 |
| Open Chinese LLM Leaderboard | 71.65 | Official Nonzero-Task Mean (%) | 97.6 |
| Open LLM Leaderboard - BBH | 63.73 | Score | 88 |
| Open Chinese LLM - GSM8K | 61.18 | Accuracy (%) | 85.8 |
| Open Korean LLM - Ko-EQ-Bench | 34.09 | EQ-Bench score | 83.8 |
| Open Korean LLM Leaderboard | 41.39 | Average Score (%) | 82.5 |
| Open Ko-LLM Leaderboard | 41.39 | Average (%) | 82.4 |
| Open Korean LLM - Ko-GSM8K (flexible extract) | 48.9 | Exact match, flexible extract (%) | 82.4 |
| Open Korean LLM - KorNAT-Harmless | 65.43 | Normalized accuracy (%) | 80.2 |
Interactive version: theaggregate.ai/model?slug=luxia-21-4b-alignment-v1-0 · How It Works · Data refreshed daily, snapshot 2026-10-20.