Translated releases Lara Mini, a 7.2B-parameter MoE research model with 600M active parameters

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, according to Unite.AI.

Key specifications

• Total parameters: 7.2 billion
• Active parameters per token: 600 million
• Training data: 6 trillion tokens
• Language coverage: English plus 44 other languages (45 in total), including all 24 official EU languages
• Training approach: trained from scratch
• Release type: research model

Architecture

Lara Mini uses a mixture-of-experts design. For each token, the model activates four of 128 routed experts, while one shared expert remains active. This sparse routing is what keeps the active parameter count at 600 million out of 7.2 billion total.

The model has 34 layers. They combine 14 Gated DeltaNet linear-attention layers with 13 sliding-window attention layers, among other layer types.

Context window, pricing and benchmark scores were not part of the information available at the time of writing.

What this means

Lara Mini pairs a small active parameter footprint (600 million per token) with broad multilingual training data, including every official EU language. Mixing Gated DeltaNet linear-attention layers with sliding-window attention is a design choice aimed at attention efficiency, and it is notable in a model trained from scratch rather than adapted from an existing checkpoint. Because it is labeled a research model, teams evaluating it should treat it as an architecture and multilingual-coverage release rather than a production service.