Translated releases Lara Mini, a 45-language AI model for non-commercial research

Translated released Lara Mini on November 12, 2026, giving researchers a new model trained across 45 languages rather than a finished translation assistant. Trained from scratch on 6 trillion tokens, it comes with weights and a technical report for non-commercial use. The release is a base model: it continues text, without instruction-following or safety tuning.

Using a fraction of the network

Lara Mini contains 7.2 billion parameters, the learned values that shape its output, but activates only 600 million for each token—a chunk of text. Its mixture-of-experts design selects four of 128 routed components for each token, alongside one always-active shared component. The smaller active count describes how much of the model participates in processing each token, not its total size.

Those components are not divided into separate teams for separate languages. Translated’s routing analysis found that languages share experts. Related languages followed similar paths early in the network and nearly identical paths in its middle layers. The authors caution that shared routing does not establish that the model represents meaning the same way across languages.

The model’s working context—the text it can handle at once—grew from 4,096 to 131,072 tokens during a separate, 100-billion-token training stage. Its training coverage includes all 24 official European Union languages, within the broader set of 45.

A narrow test of the English tradeoff

Translated tested whether substituting more English for multilingual material would improve English results. Researchers took the same checkpoint and ran two controlled, 100-billion-token final training phases, gradually lowering the learning rate to zero. In one, they replaced the 33% multilingual share with English, raising English’s share from 27% to 60%. The measured changes were uneven:

+0.3 pointsEnglish development average

Translated reported a small English gain in the controlled final-stage training comparison.

−2.7 pointsSupported-language multilingual average

The multilingual average fell when researchers substituted English for the multilingual share.

That experiment covers only the final training phase, not multilingual data’s effect throughout pretraining. The authors also say their experiments do not demonstrate that multilingual training improves general reasoning or English performance. That limits how far the results support the company’s broader argument that linguistic diversity can enrich intelligence.

Translated’s broader evaluation compared Lara Mini with 56 other open base models on 26 benchmarks. Among the 36 models using up to 2 billion active parameters, Lara Mini’s overall average was 57.0, ahead of granite-4.0-h-tiny at 53.9 and Qwen3-1.7B at 53.8. These are company-run comparisons of base models, not finished chat systems.

The overall score gives equal weight to English, multilingual understanding, mathematics and coding. Multilingual results use Lara Mini’s supported non-English languages for every model. Lara Mini ranked first in multilingual understanding and math, second in English and fifth in code within that class. Performance also varies by language; Azerbaijani and Bengali were among its lowest-scoring supported languages, where training data are limited.

Efficiency depends on workload, too. Translated measured about 1,100 generated characters per second for one user on an A100 GPU without quantization. Its reported speed advantage was strongest with one user and long prompts, and narrowed when many users shared the GPU.

Research access now, translation integration later

Access to the weights requires a request that Translated reviews individually, according to Unite.AI. The model card warns that Lara Mini can produce wrong, biased or offensive text and should not be deployed as an assistant without further training.

• Model weights use CC BY-NC-SA 4.0 for research and other non-commercial purposes; Translated offers commercial licensing separately.
• Code and tokenizer use Apache 2.0. The release does not include training data.

Lara Mini is not yet available in Lara Translate. Translated plans to use it as the foundation for offline text translation in a future app, with processing on the phone so text stays on-device. The 45-language training coverage is not a confirmed language list for that planned offline product.