Open Model WeightsNVIDIA · FIELD VERIFIED
Primary sources last checked on 21 Oct 2026. Verification freshness is tracked separately from factual change, and unknown values remain explicit rather than inferred.
FRESHNESS & VERIFICATION
A verification check does not pretend the model changed. Open Model Weights updates the verification date after a real primary-source check, while factual-update dates move only when retained source evidence changes.
CONFLICT INTELLIGENCE
+ Save to my shortlistMonitor changes →Local browser monitoring · no background email alerts
DIRECT ANSWER · FEATURED MODEL BRIEFING
Source-grounded editorial layer · reviewed 2026-10-20
NVIDIA Nemotron 3.5 Rerank 8B BF16 is a first-party, downloadable 8B-class cross-encoder for scoring the relevance of retrieved text or visual-document candidates against a query. It is not a database, embedding index or conversational model. On 20 October 2026 Open Model Weights verified its exact upstream SHA, six listed weight-related artifacts, an official LICENSE file and NVIDIA OpenMDW 1.1 licensing. The source configuration reports 262,144 positional tokens, whereas the model-card evidence includes a 32,768-token claim; supported production pair lengths require separate tests. OMW has not independently run this checkpoint, measured its ranking gains or benchmarked memory.
WHY IT MATTERS
Many RAG systems retrieve more passages than can be fed to a generator. A reranker can evaluate question-document relationships before committing limited context to downstream answers.
Visual-document relevance requires keeping image preprocessing and query instructions aligned; a reranker can add computation per candidate, making quality versus latency a core deployment tradeoff.
MODEL-SELECTION VIEW
A cross-encoder inspects a query together with each candidate to estimate match strength, which allows interactions that a separately encoded dense vector may miss. Nemotron 3.5 Rerank is therefore a second-stage model: it should receive a bounded set of passages, pages or candidate document views from an existing search component.
Unlike Nemotron 3.5 Embed, which can precompute document representations for an index, Rerank evaluates candidate pairs at query time. The distinction changes the scaling profile: serving cost follows query volume multiplied by the number and size of candidates. Quality measurement should therefore hold the initial retrieval stage constant.
NVIDIA documents the checkpoint as a cross-encoder reranker for multilingual text and visual-document retrieval, with custom-code and text-ranking signals in the official repository. Query-to-image pair construction and the supported visual processor are part of the model contract, not optional cosmetic steps.
Build an evaluation set that includes images of documents with tables, long titles, charts and mixed languages. Verify the ranking score path under the exact documented package versions. A model advertised for visual documents should not be assumed to rank every raw PDF, unprocessed video frame or external image URI without conversion.
The published config provides 262,144 positional tokens, while the source model-card claims include 32,768. Neither number alone says how many candidate pages can be jointly processed per request, because inference wrappers, image tokenization and runtime kernels may enforce stricter practical constraints.
Treat model output as a relative relevance signal, not an independently calibrated measure of correctness. Evaluate precision and recall at fixed top-k thresholds, compare with the starting ordering, and inspect error cases where a passage is topically close but factually unhelpful.
Measure a rerank-only baseline as well as the full search-to-answer path. A reranker can improve top-results quality while increasing total time per query. Record latency by query length, candidate number, document length, language and input modality, and preserve the actual first-pass retrieval order for comparison.
The original checkpoint exposes about 8.38 billion parameters in weight metadata, with BF16 source artifacts. That is sufficient to plan realistic capacity tests but not to claim peak VRAM or tokens per second. Measure on the chosen accelerator and record memory fragmentation, truncation and batch behavior under concurrency.
Even an accurately ranked source may be stale, malicious or unrelated to the user's actual authorization. Keep provenance, access-control checks and document version identifiers attached to candidates through the reranking and generation stages. A high score never grants permission to reveal protected contents.
Run adversarial cases with prompt injection inside documents and with near-duplicate passages containing conflicting dates. Evaluate whether the final assistant cites the correct version, instead of treating high rerank scores as source verification. This prevents an ordinary ranking improvement from being marketed as an agent security guarantee.
The checkpoint's first-party repository contains a LICENSE file identifying NVIDIA OpenMDW v1.1. The agreement's conditions apply to this reranker even when another encoder is published under similar terms. Review deployment and redistribution obligations for the exact pair of models you ship.
Our judgment: use the Embed model for first-pass semantic retrieval and evaluate Rerank only when reranked precision materially improves human-scored answer support at an acceptable latency. For extracting page boxes or OCR text rather than ordering candidate evidence, the distinct Nemotron 3.5 Parse parser is more appropriate.
IMPLEMENTATION & EVALUATION
Practical source-grounded evaluation plan for the exact nvidia/Nemotron-3.5-Rerank-8B-BF16 revision, not an Open Model Weights runtime or benchmark execution.
Pin source SHA, LICENSE, query/candidate input template and model-specific custom-code dependency versions.
Store exact pair inputs and scores for text and visual-document cases as reproducible fixtures.
A generic sequence-classification wrapper may score malformed pairs or ignore visual content.
Use one retrieval snapshot and a labelled query set before applying cross-encoder reranking.
Retain original order, relevant-document judgements, candidate limits and top-k metrics.
Changing the retriever makes an apparent reranking gain impossible to attribute.
Compare matched corpora and languages at candidate sizes such as five, ten and twenty, without claiming a universal optimum.
Report top-k precision, rerank latency, peak memory and unknown or failed document types.
More candidates or long page images can erase gains from improved relevance.
Carry version, citation and authorization through rerank to final generation, including injected or stale passages.
Demonstrate refusal to expose unauthorized pages and correct citation versions in a fixed audit set.
High ranking scores cannot substitute for document authorization or content verification.
Proposed tests, not OMW benchmark results. Read the official model card at the pinned source revision ↗ Inspect OMW field evidence ↗
PRIMARY & INDEPENDENT SOURCES
FAQ
No. It scores query–candidate pairs after retrieval. Use a separate encoder or search system to generate the initial candidate list.
The publisher describes visual-document reranking, but supported image processors, raw file formats and runtime versions must be reproduced for each intended input.
No. The official model card mentions a shorter context claim, and no universal practical cross-encoder pair limit has been independently measured here.
No. Scores rank candidate relevance, not whether a source is current, authorized, correct or safe for a generative assistant to follow.
Embed builds reusable semantic representations for search. Rerank evaluates a smaller returned set with query-candidate interactions and extra inference cost.
We verify the upstream repository, SHA, weight files and license evidence. Published rankings, score calibration, latency and hardware compatibility remain untested.
MODEL INTELLIGENCE · SOURCE-FIRST
Open Model Weights records Nemotron 3.5 Rerank 8B BF16 as a 8B total (publisher-declared)-parameter query and text/image candidate → relevance score open-weight model in the Nemotron 3.5 Rerank family, with NVIDIA listed as the developer. The checked source repository provides 6 recognized weight artifacts, format evidence for Safetensors and a verified context window of 262,144 tokens. Values that are not evidenced in the checked sources remain explicitly undisclosed or unknown.
The repository declares NVIDIA OpenMDW License Agreement v1.1. Commercial and non-commercial use declared by NVIDIA; OpenMDW 1.1 checkpoint-specific conditions govern. Open Model Weights reports the checked license evidence and classification; this is not legal advice.
The checked sources declare mistralai/Ministral-3-8B-Instruct-2512-BF16 as lineage evidence. Training recipe status: training-described-in-model-card; data disclosure status: described-in-model-card.
DEPLOYMENT PROFILE
These bars visualize weight-only decimal-GB estimates. They exclude KV cache, activations, runtime overhead, optimizer state and sharding.
LICENSE & USE
Commercial and non-commercial use declared by NVIDIA; OpenMDW 1.1 checkpoint-specific conditions govern
This section summarizes the checked repository evidence. It is not legal advice. Inspect license evidence ↗
LINEAGE & PROVENANCE
Open Model Weights does not infer a hidden base model or training corpus when the checked sources do not declare one.
RUNTIME SIGNALS
Declared and model-card-mentioned runtimes are discovery signals. They are not treated as Open Model Weights runtime tests unless explicitly measured.
EVIDENCE TRAIL
This is where the article remains auditable: source links on one side, the registry's observed change history on the other.
Checked Hugging Face API metadata, exact repository file list, model card and config when accessible.
Official Hugging Face repository SHA unchanged; prior field evidence retained.
Checked Hugging Face API metadata, exact repository file list, model card and config when accessible.
01 · WEIGHTS
02 · LICENSE
Classification describes the checked license terms; it is not legal advice.
License / model-card evidence ↗03 · HARDWARE
Weight-only decimal-GB estimate from total parameter count; excludes KV cache, activations, runtime overhead, optimizer state and sharding.
04 · FORMATS
“Not observed” means no matching repository artifact/name/dtype signal was found; it does not prove no third-party conversion exists.
05 · LINEAGE
06 · TRAINING ASSETS
07 · RUNTIME SUPPORT
Open Model Weights separates model-repository evidence from external runtime-release evidence. An external support statement does not change the official model repository's weight formats and does not imply an Open Model Weights inference benchmark.
08 · SOURCES
09 · VERIFICATION HISTORY
Checked Hugging Face API metadata, exact repository file list, model card and config when accessible.
Official Hugging Face repository SHA unchanged; prior field evidence retained.
Checked Hugging Face API metadata, exact repository file list, model card and config when accessible.