Lara Mini, a research model supporting 45 languages, has been introduced

Translated, which develops translation technologies, announced Lara Mini, a research model covering 45 languages, on November 12, 2026. The company trained the model from scratch: 6 trillion tokens, English plus 44 other languages. All 24 official languages of the European Union are on the list.

7.2 billion parameters, 600 million per token

Lara Mini has 7.2 billion parameters, but it uses only 600 million of them for each token. The model activates four out of 128 routed experts, while one shared expert stays on all the time. Of the 34 layers, 14 use Gated DeltaNet linear attention, 13 use sliding attention with a 512-token window, and 7 are global attention layers. This distribution is meant to limit the growing inference state seen in models that use full attention.

After training, a separate 100-billion-token stage expanded the context window from 4,096 to 131,072 tokens. On the data side, content collected from open sources was filtered into a pool of 13.8 trillion tokens, and 97 percent of the initial training tokens were the first pass over a document. English contributed 2.49 trillion tokens (41.4 percent), while the other 44 languages contributed 2.18 trillion (36.4 percent). The remaining share came from code, math, and samples resembling supervised fine-tuning.

Languages do not use separate expert groups

Routing analysis showed that languages do not have separate expert groups and instead share experts. Related languages produce similar routing at the start of the network and nearly identical routing in the middle layers. The researchers specifically note that shared routing does not mean shared semantic representation, and that the experiments did not show multilingual training improving general reasoning or English performance.

The English substitution experiment also supports that caution. The same checkpoint was trained twice with a final 100-billion-token stage. When 33 percent of the multilingual share was replaced with English, English’s share rose from 27 percent to 60 percent. Average English improvement increased by only 0.3 points, while the average across supported languages fell by 2.7 points. The experiment covers only the final stage and does not measure the multilingual share in the main pretraining.

Ahead of its class in benchmarks, clear about its limits

Translated compared the model with 56 open foundation models across 26 English, multilingual, math, and code benchmarks. Among 36 models with no more than 2 billion active parameters, Lara Mini ranks first with an average score of 57.0. granite-4.0-h-tiny scores 53.9, while Qwen3-1.7B gets 53.8. Those two models were trained on 22.5 and 36 trillion tokens respectively, while Lara Mini’s initial training used 6 trillion tokens.

The model ranks first in its class in multilingual understanding with 56.7 and in math with 59.9, second in English with 66.4, and fifth in code with 45.1. The comparison covers base models, not instruction-following versions. Quality varies across languages; because training data is limited, Azerbaijani and Bengali are among the lowest-scoring languages.

On the speed side, the non-quantized measurement for a single user on an A100 is about 1,100 generated characters per second. All models with a higher overall average generate text at least 1.8 times more slowly. The advantage is greatest with long prompts and single-user use, and narrows when many people share the GPU.

A research model, not a ready-made product

The released file is not a chat or instruction-following model: it is based on text completion and has no safety fine-tuning. The model card says it can generate incorrect, biased, or offensive text and should not be used as an assistant without additional training. The weights are provided through application review. The license is CC BY-NC-SA 4.0 for research and non-commercial use, while the code and tokenizer are under Apache 2.0. A commercial license is sold separately, and the training data is not shared.

The company sees Lara Mini as part of the Lara 4 family. According to the additional announcement on November 13, the model is not yet in the Lara Translate app. It is planned to serve as the basis for offline translation in the next version, with text processed on the phone. The figure of 45 languages describes the training scope; it has not been announced which languages will be included at launch in the app.

Source: Unite.AI