gpt-sw3-126m [bfloat16] [rev: 9272f5a996cf]: benchmark results
AI Sweden's gpt-sw3-126m, a decoder-only causal language model developed with RISE and WASP WARA for Media and Language using NeMo Megatron GPT. Its card describes Nordic-language, English and programming-code pretraining. This row preserves revision 9272f5a996cf785b8ab706a27d1e7dff1228dc70 and its explicit bfloat16 evaluation setting; matching family names do not establish checkpoint equivalence. Provider: AI Sweden. Access: Open.
Aggregate Rating 1083 ± 42, rank #10773 of 10870 rated models, from 6 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| Open LLM Leaderboard v1 - TruthfulQA MC2 | 44.03 | MC2 (%) (0-shot) | 25.7 |
| Open LLM Leaderboard v1 - GSM8K | 0.08 | Accuracy (%) (5-shot) | 9.6 |
| Open LLM Leaderboard v1 - HellaSwag | 29.54 | Normalized accuracy (%) (10-shot) | 5.3 |
| Open LLM Leaderboard v1 - WinoGrande | 50.67 | Accuracy (%) (5-shot) | 4.1 |
| Open LLM Leaderboard v1 - MMLU | 24.43 | Accuracy (%) (5-shot) | 3.8 |
| Open LLM Leaderboard v1 - ARC Challenge | 22.18 | Normalized accuracy (%) (25-shot) | 0.9 |
Interactive version: theaggregate.ai/model?slug=gpt-sw3-126m-bfloat16-rev-9272f5a996cf · How It Works · Data refreshed daily, snapshot 2026-11-12.