llama-nemotron-rerank-1b-v2

by nvidia

Lightweight 1B text reranker for production retrieval pipelines

855Kdl/month

59likes

1.2Bparams

HuggingFace Run it on your own data, free

Identifiers

Model ID

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

Feature URI

mixpeek://reranker@v1/nvidia_nemotron_rerank_1b_v2

Overview

Llama-Nemotron Rerank 1B is NVIDIA's lightweight cross-encoder reranker based on Llama-3.2-1B. At 1B parameters, it is 3.5x smaller than mistral-based rerankers while delivering competitive quality across 26 languages and 8192-token context. It supports true/false relevance scoring: the model outputs the probability that a document is relevant to a query.

On Mixpeek, this reranker slots into the second stage of retrieval pipelines: after a fast embedding-based first stage returns top-K candidates, the reranker rescores each candidate with full cross-attention between query and document, significantly improving precision.

Architecture

Cross-encoder based on Llama-3.2-1B. Outputs relevance probability via true/false token logits. 8192-token max context. 26-language support. Trained with progressive curriculum on diverse relevance data.

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: "text_embeddings",
    version: "v1",
    parameters: { model_id: "BAAI/bge-large-en-v1.5" },
  },
});

Capabilities

1B parameters: 3.5x smaller than 4B rerankers
26-language support including CJK and European languages
8192-token context for long document reranking
Cross-lingual reranking (86.83% on MLQA)
Production-grade latency on modest hardware

Use Cases on Mixpeek

Two-stage retrieval: rerank embedding search results for higher precision

Multilingual search: rerank across 26 languages with one model

Long document reranking: handle full documents up to 8K tokens

RAG pipelines: improve context quality before LLM generation

Benchmarks

Dataset	Metric	Score	Source
NQ + HotpotQA + FiQA + TechQA	Recall@5	73.64%	NVIDIA, 2026: Model Card
MLQA (cross-lingual, 7 langs)	Recall@5	86.83%	NVIDIA, 2026: Model Card
MLDR (long docs, 13 langs)	Recall@5	70.69%	NVIDIA, 2026: Model Card