gpt-sw3-6.7B-v2: benchmark results
AI Sweden's gpt-sw3-6.7b-v2, 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. The v2 card explicitly specifies a changed data distribution with more English and code and longer training than the original 6.7B release. Provider: AI Sweden. Released 2023-04-28. Access: Open.
Aggregate Rating 1181 ± 35, rank #1583 of 1635 rated models, from 48 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| EuroEval Danish NLU - Multi Wiki QA DA | 72.38 | Character F1 (%) | 57.2 |
| EuroEval Swedish NLU - Multi Wiki QA SV | 69.48 | Character F1 (%) | 56.9 |
| EuroEval Icelandic NLU - ScaLA IS | 0.69 | Matthews correlation coefficient (%) | 33 |
| EuroEval Swedish NLU - ScaLA SV | 9.94 | Matthews correlation coefficient (%) | 30.8 |
| EuroEval Icelandic Knowledge | 2.32 | Knowledge Average Score (%) | 27.8 |
| EuroEval Faroese NLU - FoSent | 5.1 | Matthews correlation coefficient (%) | 25.1 |
| EuroEval Danish Knowledge - Danish Citizen Tests | 15.59 | MCC (x100) | 22.1 |
| EuroEval Icelandic NLU - Hotter and Colder Sentiment | 7.47 | Matthews correlation coefficient (%) | 21.9 |
| EuroEval Danish Knowledge | 9.63 | Knowledge Average Score (%) | 21.8 |
| EuroEval Danish NLU | 23.36 | NLU Average Score (%) | 21.6 |
| EuroEval Norwegian Knowledge - NRK Quiz QA | 5.61 | MCC (x100) | 21.1 |
| EuroEval Danish Knowledge - Danske Talemaader | 3.68 | MCC (x100) | 20.1 |
Interactive version: theaggregate.ai/model?slug=gpt-sw3-6-7b-v2 · How It Works · Data refreshed daily, snapshot 2026-10-15.