NVIDIA: Nemotron 3.5 Rerank 8B (free)

NVIDIA: Nemotron 3.5 Rerank 8B (free) is a reranking model from NVIDIA, published under the model id nvidia/nemotron-3.5-rerank-8b:free. It belongs to the rerank category, so its job is not to generate conversational text but to score and reorder a set of candidate items against a query. The provider release date is 2026-10-21. On api14 it is offered as a free variant, which makes it straightforward to evaluate against your own retrieval data before committing to it in production.

The model accepts text and image inputs and produces rerank output. It has a context length of 32768 tokens, and the maximum output is 29491. Because it handles both text and images, it can be used in pipelines where the items being ranked are not purely textual. The output modality is rerank rather than text, so you should expect relevance scores or an ordering rather than prose answers.

Practical use cases follow directly from that shape. A common one is second stage retrieval: take the top results from a vector or keyword search, then have this model rescore them so the most relevant items rise to the top. The same pattern applies to retrieval augmented generation, where better ordering of the retrieved passages improves the quality of whatever generator model you use downstream. Because image input is supported, it can also rank mixed media candidates, for example selecting the most relevant images or documents containing images for a given query. It is also a reasonable choice for deduplication or prioritisation tasks where you have a pool of candidates and need a defensible ordering.

You call it through api14's OpenAI-compatible API using the model id nvidia/nemotron-3.5-rerank-8b:free. Because the endpoint follows the OpenAI schema, you can point an existing OpenAI client at api14's base URL and change the model field, which keeps integration work small. The model lists no supported parameters, so you should not rely on tuning options such as temperature or top_p being accepted; send your query and candidate items in the request and read the rerank output. If you need to control ordering behaviour, do it on your side by how you construct and batch the candidates.

Given the empty supported parameter list, this model is best treated as a fixed function rather than something you steer with sampling settings. If your task is text generation, summarisation or chat, a generative model will fit better, since this one returns rerank output only. The 32768 token context and 29491 maximum output bound how much candidate material you can submit in one request, so for very large candidate pools you will need to batch. It is a sensible pick when you want a free, image aware reranker and can work within those constraints.