VectaBass Research

This is the home of our open models and research. We train autoregressive (AR) and masked diffusion (MDLM) models, for language and for audio, and we publish the models, data and results behind each paper here.

UK sovereign AI is a focus. Our training data draws on UK public sources such as gov.uk, legislation.gov.uk, Parliament, the ONS and Welsh-language collections, and we train on Isambard-AI, the UK's national AI supercomputer.

Phase 1: language scaling

We trained one ladder of language models, from 100M to 1.6B parameters, three ways: as AR models, as MDLM models, and as AR models converted to MDLM. We fitted a scaling law to the ladder and froze a prediction for a 2.9B model before training it. The trained 2.9B model passed that preregistered test.

The paper and models will be released here. Nothing is published yet.

Audio

Our AR and MDLM audio models will be published here as they are ready.

What you will find here

• Models. Base checkpoints, with the training recipe and validation loss on each card.
• Datasets. Evaluation sets we built, with how they were made and what they measure.
• Papers. Linked from each model and dataset card.

VectaBass also builds IDA, an agentic data-management platform for regulated industries.

[contact email to confirm] · vectabass.com