NVIDIA: Nemotron 3.5 Embed 1B (free)

Nemotron 3.5 Embed 1B is an embedding model from NVIDIA, released by the provider on 21 October 2026. It sits in the embeddings category rather than the chat or generation category, so its job is to turn inputs into vectors rather than to produce written answers. The name reflects the family it belongs to (Nemotron 3.5), its embedding role, and its roughly 1B parameter scale. If you are building retrieval, search or clustering on top of an API, this is the kind of model you would evaluate alongside other embedding endpoints rather than instead of a chat model.

The model is multimodal on the input side. It accepts text and images, and it returns embeddings as its only output modality. That combination is what lets it embed both a text query and a document page image into the same vector space. The context window is 32768 tokens, and the maximum output is listed as 29491. NVIDIA describes it as a multimodal embedding model for multilingual retrieval over text and document page images, including tables, charts and infographics. In practice that means the model is aimed at documents where the useful information is not always in clean, extractable text.

The use cases follow directly from those capabilities. Multilingual retrieval is the headline one: you can index a corpus in more than one language and match queries against it without translating everything first. Because page images are accepted as documents, you can embed scanned pages, PDFs rendered as images, and pages whose meaning lives in tables, charts or infographics, then retrieve them by a text query. Typical applications include document search over mixed text and image collections, question answering systems that retrieve pages before a separate generation step, deduplication or clustering of documents, and recommendation or similarity features built on top of the vectors.

You call it through api14's OpenAI-compatible API using the model id "nvidia/nemotron-3.5-embed-1b:free". Because the surface is OpenAI-compatible, you point an existing embeddings client at the api14 base URL and pass that model id, so code written for other embedding providers usually needs only the base URL and model name changed. The model's supported parameters list is empty, which means there are no extra tuning parameters documented for it beyond what the embeddings endpoint itself accepts. Treat it as a straightforward embed-and-return-vectors endpoint rather than one you configure per request.

One thing to weigh is that this is an embedding model, so it will not generate text, hold a conversation or answer questions on its own. If your task is generation, reasoning or chat, a chat model is the right fit and this one is not a substitute. It is also offered under a free model id, which is useful for evaluation and low-volume work, but you should confirm current availability and any usage conditions on the model page before depending on it in production. If you need a model that both retrieves and answers, plan for a two-stage setup: embed and retrieve with this model, then pass the retrieved content to a separate generation model.