Poro 34B
Poro 34B is a Finnish and English base language model trained on text and code. The collaboration trained it from scratch using the BLOOM architecture.
Research release
Model design and training
The release reports one trillion training tokens. It uses ALiBi position encoding and a custom 128,000-token vocabulary. Training ran on 512 AMD MI250X GPUs on LUMI. Earlier training checkpoints remain repository branches; they are stages of this release, not independent model families.Poro-34B: official model card
Practical assessment
Practical guidance. Use Poro to study Finnish pretraining or to train a task-specific model. A base checkpoint predicts continuations and does not provide a ready chat interface. Its short training sequence makes passage selection important for document work.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Status | Research release |
| Access | Downloadable |
| Model identifiers | Poro-34B |
| Context window | 2,048 tokens Sequence length reported by this release's model card; this is not a guaranteed useful retrieval length. Poro-34B: official model card |
| Model size | 34.2 billion parameters.Poro-34B: official model card |
| License | Apache 2.0Poro-34B: official model card |
Related models
- Poro 34B Chat
LumiOpen · Poro
A Finnish and English instruction model adapted from Poro 34B with supervised fine-tuning and ChatML formatting.
Launch date not confirmedDownloadable - Viking 33B
LumiOpen · Viking
A 33B base model pretrained on Finnish, English, Swedish, Danish, Norwegian, Icelandic and code.
Launch date not confirmedDownloadable