Viking 33B
Viking 33B is a base model pretrained on Finnish, English, Swedish, Danish, Norwegian, Icelandic and code. It uses a LLaMA-like architecture, not a claimed adaptation of Meta's released weights.
Model design and training
56 decoder layers. The card lists intermediate training branches and a two-trillion-token target; it does not establish that the target was completed. Viking uses rotary position embeddings and FlashAttention. The downloadable base is intended for further tuning. Training checkpoints and later packaging are grouped with the same size-specific release.Viking-33B: official model card
Practical assessment
Practical guidance. Use the larger checkpoint when studying the effect of scale on the same Nordic language mix. Compare the actual training revision as well as model size. A larger base model still needs an instruction or application-specific training stage for dialogue.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Access | Downloadable |
| Model identifiers | Viking-33B |
| Context window | 4,096 tokens Sequence length reported by this release's model card; this is not a guaranteed useful retrieval length. Viking-33B: official model card |
| Model size | 33 billion parameters.Viking-33B: official model card |
| License | Apache 2.0Viking-33B: official model card |
Related models
- Viking 13B
LumiOpen · Viking
A 13B base model pretrained on Finnish, English, Swedish, Danish, Norwegian, Icelandic and code.
Launch date not confirmedDownloadable A 7B base model pretrained on Finnish, English, Swedish, Danish, Norwegian, Icelandic and code.
Launch date not confirmedDownloadable