Translated Releases Lara Mini, a 45-Language Research Model

Translated announced Lara Mini on November 12, 2026, releasing a research model trained from scratch on 6 trillion tokens spanning English and 44 other languages, including all 24 official EU languages. The 7.2-billion-parameter system activates only 600 million parameters per token.

The mixture-of-experts architecture activates four of 128 routed experts for each token, while one shared expert remains active. Its 34 layers combine 14 Gated DeltaNet linear-attention layers, 13 sliding-window attention layers with a 512-token window, and seven global-attention layers, limiting the growth of inference state relative to full attention. After the initial training run, a separate 100-billion-token stage extended the context window from 4,096 to 131,072 tokens.

Translated built the dataset from openly released sources, filtering and deduplicating them into a 13.8-trillion-token pool; 97% of the initial training tokens were the first read of a document. English supplied 2.49 trillion tokens, or 41.4%, while the 44 other languages supplied 2.18 trillion, or 36.4%. Code, math and examples formatted like supervised fine-tuning data supplied the remainder.

Testing English Substitution

For a separate test, the researchers used the same checkpoint for two controlled 100-billion-token anneals, the final training phase in which the learning rate decays to zero. Replacing the 33% multilingual share with English raised English’s share from 27% to 60%. The technical report says the English development average gained only 0.3 points, while the supported-language multilingual average lost 2.7 points.

The experiment applies only to the anneal; it does not measure the multilingual share’s effect during the main pretraining stage. A separate routing analysis found that languages share experts rather than receiving separate expert groups, with related languages showing similar routing early in the network and nearly identical routing in the middle layers. The authors caution that shared routing does not establish shared semantic representations, and that their experiments do not demonstrate that multilingual training improves general reasoning or English performance.

Benchmark Scores and Their Limits

Translated evaluated Lara Mini against 56 other open base models on 26 English, multilingual, math and code benchmarks, using common prompts, settings and answer extraction. Among the 36 models with up to 2 billion active parameters, Lara Mini averaged 57.0 overall, ahead of granite-4.0-h-tiny at 53.9 and Qwen3-1.7B at 53.8. Their reported pretraining totals were 22.5 trillion and 36 trillion tokens, respectively, compared with 6 trillion for Lara Mini’s initial run.

Translated’s results put Lara Mini first in that class on multilingual understanding with 56.7 and math with 59.9, second on English with 66.4, and fifth on code with 45.1. The overall score weights the four domain averages equally, and multilingual benchmarks are averaged over Lara Mini’s supported non-English languages for every model. The comparison covers base models rather than instruction-tuned systems. Quality also varies across languages, with Azerbaijani and Bengali among the lowest-scoring supported languages where training data are limited.

Translated measured Lara Mini at about 1,100 generated characters per second for one user on an A100 GPU without quantization; every model with a higher overall average in its speed comparison generated text at least 1.8 times more slowly. The advantage is greatest for a single user and long prompts, and narrows when many users share the GPU.

Research Access and Product Plans

The released artifact is a text-continuation base model rather than an instruction-following or chat system, and it has no safety tuning. Its model card warns that it can produce wrong, biased or offensive text and should not be deployed as an assistant without further training.

Access to the weights is gated, with each request reviewed by Translated. The weights are licensed under CC BY-NC-SA 4.0 for research and other non-commercial purposes, while the code and tokenizer use Apache 2.0. Translated offers commercial licensing separately and is releasing no training data with the model.

Translated’s research page says selected enterprise customers will also have access and describes fine-tuning for specialized sovereign applications as an intended use. Lara Mini is part of Lara 4, the foundation-model family intended to power the next generation of Lara Translate.

A follow-up published November 13, 2026 said Lara Mini was not yet available in Lara Translate. The company planned to use it as the base for offline text translation in the next Lara Translate app, with processing on the phone so text remains on the device. The 45-language figure describes training coverage; launch language support for offline translation was left unspecified.

The planned app ties the research to a concrete product test: turning a sparse research checkpoint into private, on-device translation while retaining the efficiency and multilingual performance reported in Translated’s evaluation.