Lara Mini: Translated releases a multilingual base model, but the weights are not free for products
What is Lara Mini? It is an open base model from the Italian translation company Translated, on Hugging Face since November 11, 2026. The model has 7.2 billion parameters, of which 0.6 billion are active per token, and a third of its training data is text in 44 languages other than English, Dutch among them. The weights are under CC BY-NC-SA 4.0: research and non-commercial use are allowed, building a product needs a separate agreement with Translated. Access is gated and reviewed per request.
We read the model card on Hugging Face and the data the Hugging Face API returns for the model. The model was also in the trending briefing of November 15.
What happened
Translated created the repository translated/Lara-4-Mini-Base on Hugging Face on November 11, 2026; the last change is from November 13. On X it was shared this weekend as a new release, but the weights themselves are from earlier in the week. The card points to a technical report by the Lara team and to a blog post on Translated's site.
The model in numbers
| Property | Value according to the model card |
|---|---|
| Parameters | 7.2B total, 0.6B active per token |
| Layout | 34 layers: 14 Gated DeltaNet, 13 sliding-window attention (window 512), 7 global attention |
| Experts | 128 per layer, 4 per token, plus 1 shared expert |
| Context | 131,072 tokens after the long-context stage; untested beyond that |
| Training | 6 trillion tokens, 2.18 trillion of them in 44 languages other than English |
| License | weights CC BY-NC-SA 4.0; code and tokenizer Apache 2.0 |
The tags on the card list all 24 official EU languages, along with Arabic, Hindi, Japanese and Chinese among others. Translated reports an average score of 57.0 across English, multilingual, math and code tasks, the highest of the 36 open base models with up to 2 billion active parameters it compared. According to the card the model produces about 1,100 characters per second for a single user on one A100 GPU. These figures come from the maker and have not been independently measured.
What this means
Lara Mini is a small language model that gets its speed from mixture of experts: only a small part of the network computes each token. For anyone building in languages other than English, a training mix that is one third multilingual is worth a look. Whether that yields good output in a given language needs your own test; we covered how in our piece on evaluating models in Dutch.
Two limits weigh heavily. It is a base model: it continues text, does not follow instructions and has no safety tuning. The card warns it can produce text that is wrong, biased or offensive. And the license rules out commercial use. On top of that, part of the training data is synthetic text made with Qwen and DeepSeek models; the card says the usage restrictions from those licenses carry over to Lara Mini and to models built on it. That fits the wider debate about open-model licenses: for weights, "open" often means "inspectable", not "free to use".
For researchers and for anyone who wants to fine-tune on their own languages, Lara Mini is a usable base. Building a product on it has to wait for an agreement with Translated, which says on the card that it is open to more permissive licensing.
What we did not check
We did not request access or run the model. We did not read the technical report or the blog post; the figures come from the model card. We did not review the exact terms of the Qwen and DeepSeek licenses the card refers to.
Checked on November 15, 2026.
This article was written with AI (Claude) and checked against the sources below.
Sources
- Translated, Lara-4-Mini-Base model card: huggingface.co
- Hugging Face API, translated/Lara-4-Mini-Base (creation date, license, parameters): huggingface.co


