A.X K1 (Thinking): benchmark results
SK Telecom's open-weight A.X K1 (December 2025), a 519-billion-parameter, 33-billion-active MoE trained from scratch, with hybrid Think/Non-Think reasoning control for efficient high-capacity instruction following. Provider: SK Telecom. Released 2025-12-29. Access: Open.
Aggregate Rating 1566 ± 22, rank #826 of 2170 rated models, from 38 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| Horangi 4 - Ko-AIME 2025 | 100 | Accuracy (%) | 96.2 |
| Horangi 4 - GLP - Mathematical Reasoning | 96 | Score (%) | 88.5 |
| Horangi 4 - KMMLU | 85 | Accuracy (%) | 73.6 |
| Horangi 4 - GLP - General Knowledge | 88.96 | Score (%) | 72.1 |
| Horangi 4 - HAE-RAE Bench (without Reading Comprehension) | 92.93 | Accuracy (%) | 69.7 |
| Horangi 4 - HRM8K | 92 | Accuracy (%) | 62.5 |
| Horangi 4 - Korean Hate Speech | 62 | Accuracy (%) | 58.2 |
| Horangi 4 - HAE-RAE Bench (Reading Comprehension) | 88 | Accuracy (%) | 53.4 |
| Horangi 4 - KMMLU-Pro | 69 | Accuracy (%) | 51.9 |
| SKT A.X K2 Card - KMO 2026 (1st Round) | 85 | Accuracy (%) | 50 |
| Horangi 4 - GLP - Logical Reasoning | 80 | Score (%) | 44.7 |
| Horangi 4 - Final Score | 64.87 | Score (%) | 44.2 |
Interactive version: theaggregate.ai/model?slug=a-x-k1-thinking · How It Works · Data refreshed daily, snapshot 2026-11-05.