Salamandra 7B
Salamandra 7B is a multilingual base model trained by BSC. It is the language backbone used by several later Salamandra adaptations.
Model and training
Its decoder has 32 layers and a hidden width of 4,096, with a 256,000-token vocabulary. The current card reports 12.875 trillion pretraining tokens spanning 35 European languages and code. It was trained from scratch, rather than created by serving an unchanged Llama checkpoint. FP8 and GPTQ releases are packaging variants.Salamandra 7B: official model card
Practical use
Practical guidance. This model is a useful starting point when the task needs further language or domain training. Keep base-model evaluations separate from chat evaluations. The larger network may improve some tasks over 2B, but the gain must justify memory and response time on the intended hardware.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Access | Downloadable |
| Model identifiers | salamandra-7bsalamandra-7b-base-fp8salamandra-7b-base-gptq |
| Context window | 8,192 tokens Maximum sequence length documented for these weights; the prompt and generated text share the window. Salamandra 7B: official model card |
| License | Apache License 2.0.Salamandra 7B: official model card |
| Model size | 7,768,117,248 parameters.Salamandra 7B: official model card |
Related models
- Salamandra 2B
Barcelona Supercomputing Center · Salamandra
Two-billion-parameter multilingual base language model pretrained from scratch on European-language text and code.
Launch date not confirmedDownloadable The original Salamandra 7B Instruct is BSC's first conversational adaptation of its 7B multilingual base.
Launch date not confirmedResearch releaseDownloadable