You can use generated results for commercial purposes. You own all rights to the content you create. The only restriction: make sure your prompt does not include copyrighted third-party material. You are responsible for respecting the rights to any input data.

The model converts text into vector representations that reflect meaning rather than word matches. This is used to build semantic search, RAG pipelines with knowledge bases, document clustering and deduplication, request classification, finding similar materials, and ranking search results.

The output is not text, but a fixed-dimension numerical vector for each fragment. Such vectors are stored in Qdrant, pgvector, or other storage and compared by cosine similarity. A context of up to 32,768 tokens allows encoding long documents without splitting them into small pieces.

This is an embedding model: it does not perform chain-of-thought reasoning or formulate answers, but only encodes input text into a vector. Reasoning logic is implemented by a generative model to which you pass the fragments found via vector search in a RAG scheme.

Our data does not indicate such capabilities for the model, and it is atypical for embeddings: the API response is already structured as an array of numbers in JSON. It is more convenient to attach tools and function calling to a generative model connected via the same BotHub API.

The acceptance of other modalities is not indicated in our data, so expect text input: queries, document paragraphs, product descriptions, tickets. For media, use separate models, and then vectorize their text transcripts and descriptions with this model.

The Qwen3 Embedding series was developed as multilingual, so Russian documents and queries are encoded in a common semantic space and found by meaning, not by exact word matches. You should check the quality on your corpus with a test set: label query-document pairs and measure search metrics.