Lara Mini explained: a 0.6B-active model for 45 languages

By LLMTest Lab · Nov 13, 2026 · 4 min read open-weightsmultilingualmoe
On this page

On this page

  1. What is new
  2. The numbers, and who reported them
  3. Access and the catches
  4. Where it sits
  5. Who should care

Not tested by us yet. Lara Mini ships as a gated base model for research, not as a hosted API, and our harness scores chat models over an API. Everything below comes from Translated's announcement and the Lara-4-Mini-Base model card on Hugging Face. We'll update this post when the weights or a hosted endpoint let us run it against small multilingual models on our own prompts.

Lara Mini is the first general-purpose language model from Translated, the Rome-based translation company. It is a mixture-of-experts base model with 7.2B total and 0.6B active parameters, trained from scratch on 6 trillion tokens across 45 languages. If you fine-tune small models for European or multilingual work and can live with a non-commercial license, it is worth a download request. If you want a chat model to call today, this is not it.

Lara Mini (Base)
Maker Translated
Announced 2026-11-12
Parameters 7.2B total, 0.6B active per token
Context 128k tokens
Training data 6T tokens, 45 languages including all 24 official EU languages
Architecture Decoder-only hybrid of Gated DeltaNet, sliding-window and global attention, 34 layers, 128 routed experts (top-4) plus one shared
Tuning Base model only. No instruction, chat or safety tuning
Access Gated weights on Hugging Face, reviewed per request
License Weights CC BY-NC-SA 4.0 (non-commercial); code and tokenizer Apache 2.0
Price No hosted API, so no per-token price

What is new

Most small open models are English-first with a multilingual layer added later. Translated trained the other way round. Per the card, about 2.18T of the 6T tokens are multilingual and about 2.49T are English, with code at 0.77T and math at 0.35T. Translated also says it drew on more than 20 years of translation data it has collected.

The company reports one experiment that explains the design. Replacing multilingual data with extra English gave almost no English gain and significantly reduced multilingual ability. That is a vendor claim from internal runs, and the announcement shows it only as a statement.

The architecture is built for long inputs on modest hardware. Only 7 of the 34 layers use global attention, so memory does not grow with input length in the other 27. The card lists a 128k context and says behavior beyond 131,072 tokens is untested.

The numbers, and who reported them

Every figure here is Translated's. We have not reproduced any of it.

Claim Value Source
Overall average 57.0 model card
English / multilingual / math / code 66.4 / 56.7 / 59.9 / 45.1 model card
Rank Highest average among 36 open base models with up to 2B active parameters model card
Named comparisons Qwen3-1.7B 53.8, granite-4.0-h-tiny 53.9, Llama-3.1-8B 51.2 model card
Speed About 1,100 characters per second for one user on one A100 model card
Long context RULER 76.6 at 32k and 63.7 at 128k; ONERULER 90.1 and 80.9 model card

Read the rank claim carefully. It covers base models up to 2B active parameters, and it is a weighted average chosen by the vendor. The announcement says the comparison set includes models from Google, Meta and Alibaba and that Lara Mini beats models trained on up to six times more data, but the page gives no per-benchmark scores in text, only a chart. The card also says every higher-scoring model in its comparison is at least 1.8x slower, which is a statement about the other models, not about Lara Mini's absolute speed.

Translated also released a technical report and a multilingual reasoning benchmark with 110 language variants. The announcement does not name it. That benchmark could be useful on its own for anyone who compares multilingual models.

Access and the catches

  • Gated download. You share your name and affiliation and the team reviews the request.
  • Non-commercial weights. The card says Translated is open to more permissive terms and points to research@translated.com for commercial use.
  • Third-party terms. Some training data was generated with Qwen, DeepSeek and Phi-4 models. The card says the DeepSeek agreement restricts certain uses and asks that the restriction be passed on when you distribute the model.
  • Base model. It continues text. It will not follow instructions reliably, and it has no safety tuning, so it can produce wrong, biased or offensive text.
  • Tooling. It needs transformers>=5.14 with trust_remote_code=True, and the card says not to prepend a BOS token.

Where it sits

Translated says Lara Mini is part of Lara 4, a family meant to power the next version of its Lara Translate service, and calls this model a first step. No date is given for larger models. For now the useful comparison is on paper, with the small open models in the card's table. For hosted multilingual options you can call today, see our roundup of the best LLMs for multilingual support, and for how a 7B-total, 0.6B-active design works, our mixture-of-experts explainer.

Who should care

  • Teams fine-tuning small multilingual models: the 45-language mix, 128k context and low active parameter count are aimed at you. Check the license before you build a product on it.
  • Researchers: the multilingual reasoning benchmark and the technical report are the parts to read first.
  • Anyone who needs a chat model now: skip it. Our Chariot-1-TTS-Multilingual write-up covers a different multilingual release you can use as a service.

We'll add a retest note once we can run the model against a fixed set of multilingual prompts.

Related articles

Chariot-1-TTS-Multilingual explained: Indian-language TTS
Chariot-1-TTS-Multilingual is a text-to-speech model for 11 Indian languages plus English, with 50+ voices and code-mixing. Specs, pricing and claims.
Best LLMs for multilingual support in 2026: 5 languages tested
The best LLMs for multilingual apps in 2026: how frontier and budget models handle French, German, Japanese, and Arabic, with the token-cost math.