NVIDIA: Llama Nemotron Embed VL 1B V2 (free)

Llama Nemotron Embed VL 1B V2 is an embedding model from NVIDIA, distributed on api14 under the model id nvidia/llama-nemotron-embed-vl-1b-v2:free. It belongs to the embeddings category, and its stated purpose is multimodal question-answering retrieval. It is a 1B class model, which places it among the smaller options in the Nemotron family, and it is offered here in a free variant.

The model accepts text and images as input and produces embeddings as output. It can embed documents in three forms: image only, text only, or image and text together. That flexibility is the core of its design, since retrieval systems often need to match a text question against a page image, a screenshot, a diagram, or a passage of prose. The context window is 131072 tokens, and the maximum output is 117964, so long documents can be embedded without splitting them into many fragments.

Practical uses follow from multimodal retrieval. A team building a question-answering system over scanned documents, PDFs, product images, or mixed media knowledge bases can embed each item once and then run similarity search against a text query. Because image and text can be embedded jointly, the same index can serve queries about a chart and queries about the paragraph next to it, which avoids maintaining separate pipelines for visual and textual content. It also suits deduplication, clustering, and recommendation over collections that mix screenshots with written material.

To call it through api14, use the OpenAI-compatible embeddings endpoint with the model id nvidia/llama-nemotron-embed-vl-1b-v2:free. The model lists no supported parameters, so requests should rely on the standard fields the endpoint requires rather than sampling controls such as temperature or top_p, which do not apply to embedding models. Send your input in the usual embeddings request shape and read the vector back from the response.

One thing to weigh is that this is an embedding model, not a chat or generation model, so it will not answer questions or write text. If your task is generation, reasoning, or conversation, a different model is the right choice. It is also worth checking that your retrieval stack can store and compare the vector dimensions the model returns, and that your inputs fit within the 131072 token context.