llama
7B parameters
Commercial OK
Reviewed May 2026

Salamandra 7B

Salamandra 7B is a base language model from Barcelona Supercomputing Center, pretrained on 12.875 trillion tokens across 35 European languages and code. It is not instruction-tuned — this is a raw foundation model. Apache 2.0 license means no commercial restrictions.

License: apache-2.0·Context: 8,192 tokens
BLK · VERDICT

Our verdict

OP · Fredoline Eruo|VERIFIED MAY 28, 2026
9.4/10

If you need a solid multilingual European base model to fine-tune on Spanish or other Iberian languages, Salamandra 7B is a reasonable starting point — BSC has credibility and the training data scale is serious. That said, this is not a drop-in chat model; you will need to do your own instruction tuning or look for community fine-tunes. The low download count means you are somewhat on your own for troubleshooting. Hedge: worth watching, but wait for an instruct variant unless you are actively building a fine-tune.

Why this rating

Auto-generated rating (Opus 4.7 judge, claude-opus-4-7). Overall 9.40/10. License (Apache 2.0) is explicitly verified in the HF card and commercial use is confirmed. Architecture details (7.77B params, 8192 context, llama-family decoder) match the card precisely. The description is honest and operator-grade — it correctly flags the base-model-only nature, low traction, and lack of published benchmarks. Best use case is appropriately specific (Spanish/Catalan fine-tuning base). The verdict is appropriately hedged and useful for a runlocalai reader weighing whether to invest in fine-tuning effort. Minor brand-fit deduction because base models with no instruct variant are a narrower audience, but the row handles that honestly.

Overview

Salamandra 7B is a base language model from Barcelona Supercomputing Center, pretrained on 12.875 trillion tokens across 35 European languages and code. It is not instruction-tuned — this is a raw foundation model. Apache 2.0 license means no commercial restrictions.

Strengths

  • Pretrained on 12.875 trillion tokens of curated multilingual data
  • Covers 35 European languages including Spanish and co-official Iberian languages
  • Apache 2.0 — fully commercial-friendly, no strings attached
  • 7B parameters fits comfortably in consumer VRAM (8–16 GB depending on quantization)

Weaknesses

  • Base model only — no instruction following out of the box, requires fine-tuning for chat or task use
  • 8,192 token context is unremarkable by current standards
  • Low community traction so far (885 downloads, 29 likes on HF) — limited third-party evaluations available
  • No published benchmark scores provided in source data

Quantization variants

Each quantization trades model quality for file size and VRAM. Q4_K_M is the most popular starting point.

QuantizationFile sizeVRAM required
Q4_K_M3.9 GB5 GB

Get the model

HuggingFace

Original weights

huggingface.co/BSC-LT/salamandra-7b

Source repository — direct quantization required.

Hardware that runs this

Cards with enough VRAM for at least one quantization of Salamandra 7B.

Compare alternatives

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?

5GB of VRAM is enough to run Salamandra 7B at the Q4_K_M quantization (file size 3.9 GB). Higher-quality quantizations need more.

Can I use Salamandra 7B commercially?

Yes — Salamandra 7B ships under the apache-2.0, which permits commercial use. Always read the license text before deployment.

What's the context length of Salamandra 7B?

Salamandra 7B supports a context window of 8,192 tokens (about 8K).

Source: huggingface.co/BSC-LT/salamandra-7b

Reviewed by RunLocalAI Editorial. See our editorial policy for how we research and verify model claims.

Related — keep moving

Before you buy

Verify Salamandra 7B runs on your specific hardware before committing money.