Three new AI efforts are attacking the same challenge from different angles: how can powerful systems do more with smarter design, better data, and clearer safety planning? Lara Mini brings multilingual research to a compact model, Genesis III targets high-quality scientific training data, and MIT FutureTech has joined a broad collaboration focused on mitigating AI risks.
Together, these projects point toward an AI race measured by more than parameter counts. Architecture, language coverage, training data, and risk planning are becoming just as important as scale.
Lara Mini Packs Multilingual Research Into a Compact System
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. The 45-language coverage includes all 24 official EU languages, giving the model a broad foundation across European and global text.
The system contains 7.2 billion parameters, but it activates only 600 million parameters for each token. That design uses a mixture-of-experts architecture, routing every token through four of 128 routed experts while keeping one shared expert active. The result is a model that offers a large parameter pool without using the entire system for every piece of text.
Its 34 layers combine three different processing approaches: 14 Gated DeltaNet linear-attention layers, 13 sliding-window attention layers with a 512-token window, and seven global-attention layers. This mix limits the growth of inference state compared with full attention, an important design choice for a model that must handle long inputs.
Translated added that long-input ability after the initial training run. A separate 100-billion-token stage extended Lara Mini’s context window from 4,096 to 131,072 tokens, giving the model room to process far longer documents and sequences.
Language Balance Becomes a Research Question
The training data came from openly released sources that were filtered and deduplicated into a 13.8-trillion-token pool. The initial run used 6 trillion tokens, and 97% of those tokens represented the first read of a document. English supplied 2.49 trillion tokens, or 41.4%, while the 44 other languages supplied 2.18 trillion tokens, or 36.4%.
Code, math, and examples formatted like supervised fine-tuning data supplied the remainder. That mix reflects an effort to build more than a general text model, with training material designed to support programming, mathematics, and instruction-style tasks.
Translated also tested how language balance affected performance during the final training phase. Using the same checkpoint, researchers ran two controlled 100-billion-token anneals, the stage in which the learning rate decays to zero. One version replaced the 33% multilingual share with English, raising English’s share from 27% to 60%.
The English development average gained only 0.3 points, while the supported-language multilingual average lost 2.7 points. The technical report makes a clear boundary around that result: the experiment applies only to the anneal and does not measure the multilingual share’s effect during the main pretraining stage.
A separate routing analysis found that languages share experts instead of receiving separate expert groups. Related languages showed similar routing early in the network and nearly identical routing in the middle layers, but the authors caution that shared routing does not establish shared semantic representations. Their experiments also do not demonstrate that multilingual training improves general reasoning or English performance.
Benchmarks and Data Quality Shift the Small-Model Race
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.
Those results stand out against the reported pretraining totals. Granite-4.0-h-tiny used 22.5 trillion tokens, and Qwen3-1.7B used 36 trillion, compared with 6 trillion for Lara Mini’s initial run. Translated’s results placed Lara Mini first in that class for multilingual understanding.
Another release takes a different route to stronger small models. Tether introduced Genesis III, an open synthetic STEM data set designed to train specialized AI models focused on science, technology, engineering, and mathematics. The project builds on QVAC Genesis I and Genesis II, released on October 24 and December 22, 2025, expanding the combined corpus to 191.43 billion tokens across 159.6 million documents.
The QVAC Genesis III paper was accepted and presented at the 2026 Conference on Language Modelling, or COLM. Testing showed that models trained on Genesis III performed better on science and STEM tests than models trained on comparable open-source data sets.
A 1.7-billion-parameter model trained on one part of the data set delivered valid answers in 99.45 percent of benchmark responses, while models trained on the full data set also showed significant improvements. The project’s message is direct: “In a nutshell, Genesis III is a move from quantity to quality.” The goal is to improve what a smaller model learns, not simply fill it with more data.
AI Capability Meets Risk Planning
Alongside these model and data releases, MIT FutureTech worked with the University of Queensland and more than 40 collaborators, including IBM Research and Stanford University, to develop a roadmap for organizations seeking to mitigate AI risks. The work includes MIT Sloan School of Management and the MIT Computer Science and Artificial Intelligence Laboratory among its MIT affiliations.
The MIT press release is dated Nov. 12, 2026, 09:30 AM. Its presence alongside Lara Mini and Genesis III adds a third piece to the emerging picture: AI progress depends not only on models that understand more languages or datasets that sharpen scientific performance, but also on plans that help organizations manage the risks those systems create.
The next phase of AI research is taking shape across all three fronts. Smaller models are becoming more selective about which parameters they activate, training data is being judged by what it teaches rather than how much it contains, and risk mitigation is moving into the same conversation as performance. That combination could define how useful the next generation of AI becomes.
Based on
