Salamandra 2B
Salamandra 2B is BSC's smaller multilingual base language model. It was pretrained from scratch on European-language text and code.
Model and training
The current card reports 12.875 trillion training tokens across 35 European languages and code. The decoder has 24 layers, a 2,048-wide hidden representation and a 256,000-token vocabulary. It uses rotary position embeddings, SwiGLU and RMSNorm. This base release has no conversational instruction tuning.Salamandra 2B: official model card
Practical use
Practical guidance. Use the base weights to study language adaptation or train a narrow task model. A completion prompt and a chat conversation are different inputs. For an assistant, compare an instruction-tuned release before building your own tuning process. Check performance in each intended language, including short informal text.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Access | Downloadable |
| Model identifiers | salamandra-2bsalamandra-2b-base-fp8salamandra-2b-base-gptq |
| Context window | 8,192 tokens Maximum sequence length documented for these weights; the prompt and generated text share the window. Salamandra 2B: official model card |
| License | Apache License 2.0.Salamandra 2B: official model card |
| Model size | 2,253,490,176 parameters.Salamandra 2B: official model card |
Related models
Early instruction-tuned adaptation of Salamandra 2B for prompts and conversation.
Launch date not confirmedResearch releaseDownloadable- Salamandra 7B
Barcelona Supercomputing Center · Salamandra
Seven-billion-parameter multilingual base model that forms the language backbone of later Salamandra adaptations.
Launch date not confirmedDownloadable