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 Embed 8B BF16 is a downloadable, source-verified embedding checkpoint designed to represent queries, text passages and visual documents for semantic retrieval. It is not a chat model or a cross-encoder reranker. Open Model Weights checked the first-party repository SHA, five listed weight-related artifacts and NVIDIA OpenMDW 1.1 license classification on 20 October 2026. The source configuration advertises 262,144 positional tokens, while the model card includes a 32,768-token claim; neither establishes a tested end-to-end image or text retrieval budget. No latency, retrieval accuracy or hardware capacity has been independently measured by this site.
WHY IT MATTERS
Visual-document retrieval can index information that plain text extraction misses, including page-level formatting, screenshots and embedded visual context. That broadens RAG candidate discovery but does not solve answer grounding by itself.
An embedding checkpoint and a reranking checkpoint have different roles. Embed builds search representations; the separate Nemotron 3.5 Rerank model scores already retrieved candidates.
MODEL-SELECTION VIEW
Nemotron 3.5 Embed 8B turns supported search queries and content into numerical representations for candidate discovery. Applications compare these vectors or supported model-defined representations to retrieve likely relevant items before a separate model composes an answer. The output should not be interpreted as a factual answer, a citation or a model-generated paragraph.
Compared with a cross-encoder reranker, an embedding model is generally applied once when documents enter the index and then again to each query. This design affects storage, refresh costs and latency. Use the separately published Nemotron 3.5 Rerank checkpoint only after candidate generation, where document-query interaction can improve ranking at additional compute cost.
NVIDIA lists visual-document retrieval, image embeddings, text embeddings and multilingual usage in the official repository metadata. This makes the checkpoint relevant for scanned material, page images and text passages, but supported modalities depend on the exact processor and inference interface rather than on model tags alone.
Before adopting it for a multilingual corpus, sample each target language and document type. A model-card language list is a publisher declaration, not proof that a low-resource language, scanned chart, small screenshot or mixed-script page meets your retrieval-accuracy threshold. Record both OCR preprocessing and original-image routes when comparing them.
The checked model configuration exposes 262,144 positional tokens, whereas the original model-card evidence includes a 32,768-token claim. These fields describe different evidence surfaces. Open Model Weights does not assert that 262K text tokens can be passed through every supported image/text wrapper with stable relevance or predictable memory.
Set explicit chunk lengths and page limits before building a large index. Measure missed-answer cases as chunks become longer, and check how page images are resized or segmented. The cost of an oversized embedding pass includes preprocessing time, memory and potential truncation, not merely the number printed in config.json.
Construct a fixed set of queries with human-checked relevant passages, including negative queries and near duplicates. Compare the model against a lexical baseline and an existing dense retriever using the same corpus, chunking policy, evaluation metrics and allowed index size. Report recall at a fixed candidate cutoff instead of an unexplained vendor score.
Keep index-generation revisions explicit. Changing image preprocessing, prompt templates, vector normalization or checkpoint revision can make old document embeddings incompatible with new query outputs. Version the entire collection, not just the model name, and use a staged reindex plan before production changes.
The publisher labels this checkpoint as an 8B-class BF16 embedding model; the exact repository's Safetensors metadata exposes roughly 8.38 billion stored parameters. Neither figure directly tells an operator peak GPU memory, request concurrency or CPU throughput, and they should not be substituted for a measured deployment.
Use the first-party model-card instructions to test Sentence Transformers, Transformers or another explicitly supported runtime. Instrument full memory, image preprocessing, batch performance and error logs on the intended accelerator. An available repository revision is evidence of downloadable weights, not evidence that any arbitrary server can run them.
NVIDIA declares OpenMDW License Agreement v1.1 for this checkpoint. A metadata license label and the terms governing weights, redistribution and downstream applications must be read together. The underlying model-card lineage mentions a Ministral 3 8B base, which is not evidence that all base-model licenses or architectural features carry over unchanged.
Our assessment: start with Embed if the problem is finding relevant documents efficiently, and add the separate Rerank model only when controlled evaluation demonstrates a net ranking benefit. For OCR-intensive extraction where structured page coordinates are required, compare the purpose-built Nemotron 3.5 Parse model instead.
IMPLEMENTATION & EVALUATION
Practical source-grounded evaluation plan for the exact nvidia/Nemotron-3.5-Embed-8B-BF16 revision, not an Open Model Weights runtime or benchmark execution.
Capture the exact checkpoint SHA, processor files, embedding settings and OpenMDW agreement before index creation.
Store a reproducible input/output sample, revision manifest, dimensions and applicable license notes.
A changed preprocessing library can silently alter all indexed vectors.
Select representative queries and document pages with human-checked relevance labels for each target modality.
Keep positive, near-negative and no-answer queries per language with a stable evaluation cutoff.
Publisher language labels do not validate your real screenshots or uncommon scripts.
Verify supported query/document encoders, normalization and vector storage format before comparing approximate search methods.
Record vector shapes, duplicate behavior, index settings, recall and end-to-end error paths.
Query and document transformations can mismatch despite using the same checkpoint name.
Benchmark batch indexing, incremental refresh and query serving with the same hardware and explicit preprocessing limits.
Log peak memory, per-page preprocessing time, retrieval accuracy and rollback criteria.
A source-configured context size and weight-only arithmetic do not predict production GPU memory.
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 is intended to create retrieval representations for search queries and documents. A separate generation model is needed to formulate user-facing answers.
NVIDIA describes visual-document retrieval and image embeddings, but each deployment must verify the exact image processor, resolution, input format and indexing contract.
A source configuration contains 262,144 positional tokens while the model card mentions 32,768. The actual supported and useful retrieval length was not independently established.
No. An embedder produces initial candidates while a reranker scores candidate pairs; their computational and relevance roles are different in a search pipeline.
NVIDIA identifies OpenMDW Agreement version 1.1, with conditions specific to these weights. Review the complete terms and deployment dependencies before redistribution.
No. We verified publisher files, SHA, source fields and licensing evidence, but have not executed this model or measured hardware memory and retrieval quality.
MODEL INTELLIGENCE · SOURCE-FIRST
Open Model Weights records Nemotron 3.5 Embed 8B BF16 as a 8B total (publisher-declared)-parameter text or document images → dense retrieval embedding open-weight model in the Nemotron 3.5 Embed family, with NVIDIA listed as the developer. The checked source repository provides 5 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.
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.
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.