Salamandra 7B Instruct
Salamandra 7B Instruct is an Apache 2.0 instruction-tuned model from Barcelona Supercomputing Center, pretrained from scratch on 12.875 trillion tokens across 35 European languages and code. It uses ChatML formatting and supports general conversational tasks. No RLHF alignment means content safety guardrails are minimal.
If you need a commercially usable, Spanish-strong base for chat that you control entirely, Salamandra 7B Instruct is a reasonable starting point. The massive pretraining corpus is genuinely impressive for a regional open model, and the Apache 2.0 license removes legal headaches. That said, the lack of RLHF is a real operational risk — don't deploy this customer-facing without your own safety layer. Hedge: worth testing for Spanish-language internal tools, but benchmark it against Mistral 7B Instruct before committing.
›Why this rating
Auto-generated rating (Opus 4.7 judge, claude-opus-4-7). Overall 9.50/10. License (Apache 2.0), parameter count (~7.77B), context (8192), and vendor (BSC-LT) all match the model card verbatim. The description honestly flags the proof-of-concept status and lack of RLHF — exactly the operator-grade framing runlocalai readers need. Use case is appropriately scoped to Spanish/European multilingual chat prototyping, and the verdict gives a clear hedge against Mistral 7B Instruct. Minor nit: the row claims '35 European languages' which matches the card, but the language list includes 'code' which isn't a language — the description correctly says '35 European languages and code', so it's accurate. Solid publication-ready row.
Overview
Salamandra 7B Instruct is an Apache 2.0 instruction-tuned model from Barcelona Supercomputing Center, pretrained from scratch on 12.875 trillion tokens across 35 European languages and code. It uses ChatML formatting and supports general conversational tasks. No RLHF alignment means content safety guardrails are minimal.
Strengths
- Pretrained on 12.875T tokens with strong coverage of Spanish and other European languages
- Apache 2.0 — fully commercial-friendly, no strings attached
- Flash attention + grouped query attention keeps inference efficient at this size
- Instruction-tuned for chat via standard ChatML template
Weaknesses
- No RLHF — more likely to produce harmful or off-policy outputs than safety-tuned peers
- BSC describes this as a proof-of-concept; expect rough edges in quality
- 8K context is workable but unremarkable by current standards
- 79 HF likes and 126K downloads suggest limited community validation so far
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 | 3.9 GB | 5 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 7B 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 7B Instruct?
Can I use Salamandra 7B Instruct commercially?
What's the context length of Salamandra 7B Instruct?
Source: huggingface.co/BSC-LT/salamandra-7b-instruct
Reviewed by RunLocalAI Editorial. See our editorial policy for how we research and verify model claims.
Related — keep moving
Verify Salamandra 7B Instruct runs on your specific hardware before committing money.