Pleias

Pleias takes a deliberately contrarian position in the foundation-model market. While frontier labs chase scale and raw capability, Pleias optimizes for three constraints that matter intensely for regulated and information-intensive sectors: small model size (so inference runs cheaply on commodity hardware), full data provenance (every training token comes from clearly licensed or public-domain sources), and European sovereignty (data, infrastructure, and corporate control sit within the EU jurisdiction).

The Pleias family includes very small models — under 1B parameters — that punch above their weight on specialized retrieval-augmented generation tasks where a tightly scoped model with a strong retrieval layer often outperforms a giant general-purpose model. Common use cases are document-heavy public-sector workflows, multilingual customer support across European languages, regulatory reporting, and information extraction over highly structured corpora. The small footprint means a model can run on a single CPU or a low-end GPU, which dramatically reduces total cost of ownership and removes the dependency on hyperscale cloud GPUs.

A key selling point is the Common Corpus, a fully open and licensed pre-training dataset the Pleias team helped construct, which means their models can be deployed in environments where 'we cannot prove what is in the training data' is a blocker — government procurement, scientific publishing, legal review. Models are released openly under permissive licenses so customers can self-host, audit, and fine-tune without vendor lock-in.

Pleias positions its commercial offering around enterprise support, custom fine-tuning, and integration services for organizations that want the open models with a vendor relationship. Pricing depends on the engagement; open-weight model downloads are free. The lab is especially relevant for European public-sector and regulated-industry buyers evaluating sovereign AI options.