How they use Nebius Token Factory

Custodio Legal uses Nebius Token Factory for text generation and document embeddings. Its subprocessor register identifies GLM-5.3-Flash for text generation and Qwen3-Embedding-8B for embeddings. The embedding model runs in Finland and supports semantic search for questions answered with document citations. Its AI features also include document drafting and case status with extracted facts.

Nebius replaced Voyage AI for embeddings on September 24, 2026. Custodio reports that organization-level zero data retention has been enabled since September 4, 2026. Its compliance disclosure places text generation and embeddings in the same Nebius organization under that setting. Custodio attributes the retention protections to Nebius’s published terms and guidance. It says it has not independently audited that guarantee. The setting applies prospectively and does not cover submissions before September 4, 2026. Custodio does not claim coverage for abuse-detection records or streaming responses.

When Q&A is enabled, document text is embedded in chunks after OCR. Answers use retrieved passages, matter context, relevant firm memory and recent conversation. Drafting sends matter context rather than attached documents. Nebius also processes document excerpts to extract parties, dates, amounts, identifiers, obligations and deadlines. Automatic matter reevaluation uses those facts and case activity rather than full document text. Firms can disable indexing, fact extraction and reevaluation separately.