K2S3-v0.1: benchmark results
K2S3 release published by Changgil: a depth-expanded Mistral-7B-v0.1 derivative configured with 64 decoder layers, with added Korean tokenizer vocabulary and supervised fine-tuning on alpaca-gpt4-data and OpenOrca samples. It remains distinct from the 32-layer and 48-layer releases. Provider: Changgil. Access: Open.
Aggregate Rating 1378 ± 30, rank #6911 of 10916 rated models, from 12 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| Open LLM Leaderboard v1 - GSM8K | 61.71 | Accuracy (%) (5-shot) | 74.5 |
| Open LLM Leaderboard v1 - ARC Challenge | 60.32 | Normalized accuracy (%) (25-shot) | 47.1 |
| Open LLM Leaderboard v1 - WinoGrande | 76.09 | Accuracy (%) (5-shot) | 45.9 |
| Open LLM Leaderboard v1 - MMLU | 59.3 | Accuracy (%) (5-shot) | 44.9 |
| Open LLM Leaderboard - MuSR | 40.14 | Score | 42.9 |
| Open LLM Leaderboard v1 - HellaSwag | 79.65 | Normalized accuracy (%) (10-shot) | 38.7 |
| Open LLM Leaderboard v1 - TruthfulQA MC2 | 47.81 | MC2 (%) (0-shot) | 38.7 |
| Open LLM Leaderboard - BBH | 46.55 | Score | 37.6 |
| Open LLM Leaderboard - IFEval | 32.77 | Score | 30.6 |
| Open LLM Leaderboard - MMLU-Pro | 25.62 | Score | 26.4 |
| Open LLM Leaderboard - MATH Level 5 | 4.61 | Score | 26 |
| Open LLM Leaderboard - GPQA | 26.43 | Score | 18.3 |
Interactive version: theaggregate.ai/model?slug=k2s3-v0-1 · How It Works · Data refreshed daily, snapshot 2026-11-08.