Salamandra 2B
Salamandra 2B is a base-only transformer trained from scratch by Barcelona Supercomputing Center on 12.875 trillion tokens across 35 European languages and code. At 2.25B parameters and an 8192-token context window, it is one of the few models with serious native Spanish and co-official Iberian language coverage built in from pretraining. It is not instruction-tuned — you will need to fine-tune it before it is useful in a product.
If you need a clean, commercially licensed base model with genuine Spanish-region language depth to fine-tune on your own data, Salamandra 2B is a reasonable starting point. BSC-LT built this from scratch rather than adapting an English-first model, which matters for Iberian language quality at the token level. That said, this is strictly a base model — do not deploy it raw. If you want something you can run today without fine-tuning work, skip this and look at an instruction-tuned alternative.
›Why this rating
Auto-generated rating (Opus 4.7 judge, claude-opus-4-7). Overall 9.40/10. License (Apache 2.0) is explicitly stated in the model card and correctly marked commercial-friendly. Metadata is precisely verified: 2,253,490,176 params rounds to 2.25B, context 8192, family llama (architecture matches), vendor BSC-LT all check out. Editorial voice is honest and operator-grade — explicitly flags base-only and warns against raw deployment. Use case is appropriately sharp (Iberian language fine-tuning). Brand fit is solid for the European/multilingual local-AI niche, slightly narrower than mainstream English models but legitimate. Weaknesses honestly note the adoption gap.
Overview
Salamandra 2B is a base-only transformer trained from scratch by Barcelona Supercomputing Center on 12.875 trillion tokens across 35 European languages and code. At 2.25B parameters and an 8192-token context window, it is one of the few models with serious native Spanish and co-official Iberian language coverage built in from pretraining. It is not instruction-tuned — you will need to fine-tune it before it is useful in a product.
Strengths
- Strong Spanish and European language coverage from pretraining — not retrofitted
- 12.875T token training corpus, large for a 2B-class model
- 8192-token context window is generous at this parameter count
- Apache 2.0 license, fully commercial-friendly
Weaknesses
- Base model only — requires fine-tuning before any instruction-following use
- 2.25B parameters will underperform larger models on complex reasoning or generation tasks
- No multilingual coverage outside Europe
- Low adoption so far (2,120 downloads, 25 likes) — limited community troubleshooting resources
Quantization variants
Each quantization trades model quality for file size and VRAM. Q4_K_M is the most popular starting point.
| Quantization | File size | VRAM required |
|---|---|---|
| Q4_K_M | 1.2 GB | 2 GB |
Get the model
HuggingFace
Original weights
Source repository — direct quantization required.
Hardware that runs this
Cards with enough VRAM for at least one quantization of Salamandra 2B.
Models worth comparing
Same parameter band, plus what's one tier above and below — so you can decide what actually fits your hardware.
Frequently asked
What's the minimum VRAM to run Salamandra 2B?
Can I use Salamandra 2B commercially?
What's the context length of Salamandra 2B?
Source: huggingface.co/BSC-LT/salamandra-2b
Reviewed by RunLocalAI Editorial. See our editorial policy for how we research and verify model claims.
Related — keep moving
Verify Salamandra 2B runs on your specific hardware before committing money.