NVIDIA: Nemotron 3.5 Embed 8B (free)

NVIDIA: Nemotron 3.5 Embed 8B (free) is an embedding model from NVIDIA, published under the model id nvidia/nemotron-3.5-embed-8b:free on api14. It sits in the embeddings category, so it is built to turn inputs into vector representations rather than to generate chat replies. The provider release date is 2026-10-21. If you are choosing a model for a retrieval or similarity pipeline, this is the kind of model you would evaluate against your own data rather than a general purpose text generator.

The model is multimodal on the input side: it accepts text and images. On the output side it produces embeddings only. Its context length is 32768 tokens, and the maximum output is 29491. NVIDIA describes it as a multimodal embedding model for multilingual retrieval over text and document page images, including tables, charts and infographics. Queries are text, while documents can be text or page images. That combination is the core of what the model does: it maps a text query and a document, in either text or image form, into a shared embedding space so they can be compared.

The practical use cases follow directly from those capabilities. Multilingual retrieval is the headline one, so search across content in more than one language is a natural fit. Because document pages can be supplied as images, you can index scanned pages, PDFs and reports without a separate OCR step, and still retrieve tables, charts and infographics that plain text extraction tends to lose. Text to text retrieval over a corpus is also supported. Typical setups include semantic search, question answering over a document store, deduplication and clustering, and recommendation or ranking where you need a similarity score between a query and a candidate document.

You call it through api14's OpenAI-compatible API using the model id nvidia/nemotron-3.5-embed-8b:free. Because the interface follows the OpenAI embeddings shape, you send your input and read the vector back from the response, which keeps client code portable if you later swap models. The model lists no supported parameters, so there are no extra sampling or generation options to set; keep requests to the input itself and the model id. Note that this is an embedding endpoint, not a chat completion endpoint, so chat specific fields do not apply.

One limit worth planning around is the context length of 32768 tokens per request, which caps how much text or how large a page image you can embed at once; longer documents need chunking. Since the model returns embeddings rather than generated text, it is the wrong choice when you need summarisation, extraction or conversational answers. For those tasks, a text generation model fits better, and you would use this model alongside it as the retrieval layer.