NVIDIA: Nemotron 3 Embed 1B (free)

Nemotron 3 Embed 1B is NVIDIA’s text embedding family for multilingual and cross-language retrieval and semantic similarity. It transforms a question or passage into a dense vector which a search system can compare with other text representations. The manufacturer evaluates the documented 1B releases across thirty-four language and language-variant settings, including English, Arabic, Chinese, Russian, Japanese, Korean and Hinglish.

Encoder and representation

The maker’s model card describes a pruned, bidirectional transformer derived from Ministral 3. It has approximately 1.14 billion parameters and averages the token-level representations to obtain one embedding per text. The native output is a 2048-dimensional floating-point vector. This is a retrieval representation rather than a written answer to the question being embedded.

The manufacturer also supports retaining the first part of that representation, for example 1024 or 512 dimensions. The shortened vector must be L2-normalised again after slicing. Keep the selected representation length and normalisation consistent across the query and corpus sides of a search task. A vector’s length and the input sequence length are separate properties.

Query and passage preparation

The documented retrieval convention uses “query:” for a question and “passage:” for document text. The producer’s examples apply those roles before encoding and compare normalised vectors with a dot product, equivalent to cosine similarity. Use the preparation expected by the selected workflow instead of mixing a raw corpus representation with a differently prepared query.

A practical collection can contain technical explanations, policies or code-related documentation. A question about how a function handles a particular case should retrieve the source passage or code excerpt that describes that behaviour. Preserve identifiers, headings and relevant conditions with the actual passage text. The vector makes the candidate searchable; the retained source is what a reader or a later RAG answer stage uses.

Longer source material

The documented maximum sequence length is 32,768 tokens. Choose passage boundaries deliberately even when the model can accommodate substantial text. A complete section can preserve a definition and its exceptions; a focused excerpt can make a specific answer easier to inspect. Inputs beyond the model window need chunking or truncation, so the omitted tail should not be treated as represented content.

Keep each embedded item linked to its parent document and version. For a report, retain the section title and the explanation around the conclusion the question concerns. For a code search, keep the function name, relevant interface and behaviour beside the excerpt. This provides a concrete source artifact when the retrieval stage returns a candidate.

Family editions and result use

NVIDIA documents BF16 and NVFP4 versions of the 1B family. Its quantised model card says those two releases share an embedding space and recommends checking retrieval quality on representative material when changing versions. The declared common representation is a manufacturer statement about that named pair, not a general rule for every model with 2048 output values.

The useful result is a query vector or an indexed source collection with its model, dimension choice and preparation recorded. Return the candidate text with its identifier and examine whether it supplies the required information. The manufacturer explicitly notes that a retrieval model does not guarantee the correct passage for every question; a later reranking or answer stage works with those candidates as its own task.