llama-nemotron-rerank-vl-1b-v2

by nvidia

Visual multimodal reranker that rescores text, image, and document results

100Kdl/month

55likes

1.7Bparams

HuggingFace Run on your data

Identifiers

Model ID

nvidia/llama-nemotron-rerank-vl-1b-v2

Feature URI

mixpeek://reranker@v1/nvidia_nemotron_rerank_vl_1b_v2

Overview

Llama-Nemotron Rerank VL extends NVIDIA's reranking to visual modalities. Built on SigLIP 2 (400M vision) + Llama 3.2 (1B language), it reranks results that contain images, document pages, or mixed text-image content. It pairs naturally with nvidia/llama-nemotron-embed-vl-1b-v2 for a complete two-stage multimodal retrieval pipeline.

On Mixpeek, the visual reranker improves precision for document retrieval, product search, and any pipeline where visual content matters. After embedding-based retrieval returns top candidates, the reranker applies full cross-attention between the query and each candidate's visual+textual content.

Architecture

SigLIP 2 (400M) vision encoder + Llama 3.2 (1B) language model. 1.7B total params. Processes text-only, image-only, and mixed text+image inputs. Outputs relevance probability via cross-encoder scoring.

Mixpeek SDK Integration

import { Mixpeek } from "mixpeek";

const mx = new Mixpeek({ apiKey: "API_KEY" });

// Managed: create a collection over a bucket; Mixpeek runs this model's extractor
const collection = await mx.collections.create({
  namespace_id: "my-namespace",
  collection_name: "my-collection",
  source: { type: "bucket", bucket_ids: ["bkt_your_bucket"] },
  feature_extractor: {
    feature_extractor_name: "image_embedding",
    version: "v1",
    parameters: { model_id: "nvidia/llama-nemotron-embed-vl-1b-v2" },
  },
});

Capabilities

Multimodal reranking: text, image, and mixed document results
1.7B parameters: efficient for production deployment
Pairs with nemotron-embed-vl for two-stage multimodal retrieval
+6-7% Recall@5 improvement over embedding-only retrieval
Evaluates on ViDoRe V1/V2/V3 document benchmarks

Use Cases on Mixpeek

Document retrieval: rerank scanned pages by visual+textual relevance

Product search: rerank product images by visual similarity to query

Mixed-media search: rerank results combining text and visual content

Visual QA pipelines: improve retrieval precision before VLM generation

Benchmarks

Dataset	Metric	Score	Source
Multimodal (text query)	Recall@5 improvement	+7.2%	NVIDIA, 2025: arxiv,2501.14818
Multimodal (image query)	Recall@5 improvement	+6.9%	NVIDIA, 2025: arxiv,2501.14818
Multimodal (image+text)	Recall@5 improvement	+6.0%	NVIDIA, 2025: arxiv,2501.14818