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

NVIDIA: Llama Nemotron Rerank VL 1B V2 (free) is a multimodal reranking model from NVIDIA. It belongs to the rerank category, and its job is narrower than that of a chat or generation model: it evaluates how relevant a document is to a user query. The model is a 1.7B parameter system, and the version served here is the V2 release. NVIDIA built it specifically for the ranking stage of retrieval pipelines, including pipelines that work with images as well as text.

On input, the model accepts text and images. On output, it produces rerank results rather than generated prose. That means you give it a query alongside candidate documents, which can be document images or text, and it returns relevance judgements you can use to order those candidates. The context window is 10,240 tokens, with a maximum output of 9,216. Because both the input and output budgets sit in that range, the practical constraint is how much query and document content you pack into a single scoring call.

The obvious use case is vision RAG. If your retrieval system indexes scanned pages, screenshots, slide decks, PDFs rendered as images, or any corpus where layout and visuals carry meaning, a text-only reranker will miss part of the signal. This model scores document images against a query directly, so it can sit between your first-stage retriever and the generation step, reordering candidates so the most relevant ones reach the context window of whichever model writes the final answer. Text-only reranking still works too, so mixed corpora are fine.

You call it through api14's OpenAI-compatible API using the model id "nvidia/llama-nemotron-rerank-vl-1b-v2:free". Because the interface is OpenAI-compatible, you point your existing client at api14's base URL and pass that id in the model field. Note that this model lists no supported parameters beyond the basics, so there are no extra knobs such as sampling controls or custom response formatting to configure for it. Keep your request to the standard message content plus the model id.

That empty parameter list is worth planning around. If your application expects to tune temperature, top_p, or similar settings, this is not the model for that; a generative model would fit better there. Likewise, if your task is producing text rather than ordering candidates, a reranker is the wrong tool. Where it fits is the ranking step, especially when your documents are images and you want relevance scoring that accounts for them.