Salamandra 2B Instruct
Salamandra 2B Instruct is a transformer model from BSC pretrained from scratch on 12.875 trillion tokens across 35 European languages and code. The instruct variant is fine-tuned for instruction following using the ChatML template. Licensed Apache 2.0, so commercial use is clear.
If you need a small, commercially safe model with genuine Spanish and broader European language coverage, Salamandra 2B is worth a look — BSC's training data budget is serious for this size class. The lack of RLHF alignment is a real concern for anything user-facing; plan to add your own output filtering. Thin community adoption means you're partly in the dark on real-world quirks. Hedge: worth testing for internal or low-stakes Spanish-language tasks, but don't deploy to end users without guardrails.
›Why this rating
Auto-generated rating (Opus 4.7 judge, claude-opus-4-7). Overall 9.40/10. License is explicitly Apache 2.0 on the card and correctly flagged commercial-OK. Metadata (2B params, 8192 context, llama family, BSC vendor) matches the card exactly. Description is honest and operator-grade: it surfaces the RLHF warning verbatim from the card, acknowledges thin adoption, and gives a hedged verdict. Best use case is specific (Spanish/European multilingual instruction following on low-VRAM). The only minor drag is brand fit — it's a niche European-language model with limited GGUF ecosystem mention — but it's still a clear local-deployable option for the right reader.
Overview
Salamandra 2B Instruct is a transformer model from BSC pretrained from scratch on 12.875 trillion tokens across 35 European languages and code. The instruct variant is fine-tuned for instruction following using the ChatML template. Licensed Apache 2.0, so commercial use is clear.
Strengths
- Pretrained on 12.875T tokens — large data budget for a 2B model
- Covers 35 European languages including Spanish, with code support
- 8,192 token context window
- Apache 2.0 — no license friction for commercial projects
Weaknesses
- No RLHF alignment — outputs are not safety-tuned, needs filtering in production
- 2B parameters means it will lose to larger models on complex reasoning tasks
- Low community traction so far (3,164 downloads, 28 likes) — limited real-world feedback
- BSC describe it as still being actively improved; treat as early-stage
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 | 1.1 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 Instruct.
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 Instruct?
Can I use Salamandra 2B Instruct commercially?
What's the context length of Salamandra 2B Instruct?
Source: huggingface.co/BSC-LT/salamandra-2b-instruct
Reviewed by RunLocalAI Editorial. See our editorial policy for how we research and verify model claims.
Related — keep moving
Verify Salamandra 2B Instruct runs on your specific hardware before committing money.