What is Nemotron 3.5 Embed (1B and 8B)?
Nemotron 3.5 Embed is a pair of open embedding models from NVIDIA, released on 20 October 2026 under the permissive OpenMDW-1.1 licence: a 1B model (1.55 billion parameters) and an 8B model (8.38 billion). An embedding model turns a query or document into a list of numbers so a search system can find the closest matches. These embed text, an image of a document page, or both together, for search over PDFs with tables, charts and infographics; queries are text only.
Both are built on Mistral's Ministral 3 models with a Pixtral image encoder added to NVIDIA's text-only Nemotron 3 Embed from July 2026. NVIDIA froze the text part during training, so text vectors match Nemotron 3 Embed and an existing text index does not need rebuilding. The 1B outputs 2,048-dimension vectors that can be cut to 1,024 or 512, reads up to 32,768 tokens, and also ships in a 4-bit NVFP4 build for Blackwell GPUs.
Is Nemotron 3.5 Embed (1B and 8B) worth it?
Nemotron 3.5 Embed is now the strongest open option for searching business documents, where tables and charts carry as much meaning as the prose. On the ViDoRe v3.1 visual document retrieval leaderboard, as of 20 October 2026, the 8B ranked first at 61.17 nDCG@10 and the 1B scored 58.38, keeping 95.4% of the larger model's score. In NVIDIA's table on ViDoRe V3 page images, the 1B scores 58.02 against 52.89 for Qwen3-VL-Embedding-2B, and the 8B 60.39 against 58.29 for Qwen3-VL-Embedding-8B.
The text side is inherited and already proven. NVIDIA reports its text-only Nemotron 3 Embed 8B first on the multilingual RTEB leaderboard in July 2026 at 78.5, with the 1B at 72.4 and 71.0 on MMTEB retrieval. Because 3.5 keeps those text vectors unchanged, a team on Nemotron 3 Embed can add page images to the same index without re-embedding its text. That is a rare courtesy in a field where every version bump usually means a full rebuild.
Cost of entry is close to zero. Both sizes are free on OpenRouter for now, the licence permits commercial use, and the 1B in NVFP4 runs 27.1% faster than BF16 on a GB200 with 99.85% of its ViDoRe V3 accuracy, in NVIDIA's test. NVIDIA's NIM container served 6,031 query requests a second on an RTX PRO 6000 Blackwell, about 2.9 times vLLM.
The catch: every number above is NVIDIA's, and the software is narrow. Supported hardware is NVIDIA Ampere, Hopper and Blackwell on Linux, image input needs Transformers 5.20.0 because 5.17.0 to 5.19.0 can return wrong results with FlashAttention 2, and page images count against the same 32,768-token budget as text. Text was evaluated in 34 languages and images in 18, fewer than the 100-plus some rivals claim.
Our take: for a RAG system over PDFs, slides and scanned reports, Nemotron 3.5 Embed 1B is a sensible default today; the 8B earns its extra GPU only where a few points of recall are worth real money.
Where it earns its place
A search engineer at an equity research firm
Find the right chart or table across thousands of analyst PDFs
Embed each page as image plus extracted text with the 1B, store 2,048-dimension vectors (or 1,024 to halve storage) and query in plain text. NVIDIA's ViDoRe V3 image-plus-text score for the 1B is 58.82.
A platform team already running Nemotron 3 Embed for text search
Add scanned contracts and slide decks without re-indexing millions of text chunks
Swap the model to Nemotron 3.5 Embed 1B and embed only the new page images; NVIDIA kept the text weights frozen, so existing text vectors stay valid.
A two-person startup prototyping an AI agent with memory
Test a strong embedder before committing to GPUs
Call nvidia/nemotron-3.5-embed-1b:free on OpenRouter during prototyping, then move to vLLM or the NIM container on a single NVIDIA GPU once traffic justifies it.
Before you commit
- Setup
- pip install transformers 5.20.0 and sentence-transformers, load nvidia/Nemotron-3.5-Embed-1B-BF16 and call encode_query and encode_document, which add the query and passage prefixes; or vllm serve with vLLM 0.32.0.
- Integrations
- Transformers, Sentence Transformers, vLLM (including a /v2/embed retrieval endpoint), NVIDIA NIM; OpenRouter, Baseten, DeepInfra and FriendliAI; Elastic support announced as coming soon.
- Data and privacy
- OpenMDW-1.1 allows commercial use. Self-hosted, documents stay on your hardware. Training used public, commercially permissible datasets and synthetic data from models including Nemotron, Qwen, Gemma and gpt-oss.
- Cost at scale
- Weights are free and OpenRouter currently charges nothing; self-hosted, NVIDIA measured the 1B NVFP4 at 128.8 documents a second on a GB200 and NIM at 6,031 query requests a second on an RTX PRO 6000.
- Limits
- 32,768 tokens per input, shared between text and image tokens; images default to a 1,120-pixel longest edge, up to 1,540; text queries only; NVIDIA GPUs on Linux.
What’s good
- 8B first on ViDoRe v3.1 at 61.17 nDCG@10 (20 October 2026)
- Embeds text, page images or both into one vector space
- Text vectors identical to Nemotron 3 Embed, so text indexes carry over
- Permissive OpenMDW-1.1 licence; free on OpenRouter
- NVFP4 build 27.1% faster on Blackwell
What’s not
- All benchmarks are NVIDIA's own
- Linux and NVIDIA GPUs only
- Text evaluated in 34 languages, images in 18
- Queries must be text; image queries are not supported
- Version-sensitive: image input breaks on Transformers 5.17.0 to 5.19.0 with FlashAttention 2
Best for
Teams building search or RAG over PDFs, slides and reports who want one open model for both text and page images.
Nemotron 3.5 Embed (1B and 8B) alternatives
Nemotron 3.5 Embed (1B and 8B): common questions
- Is Nemotron 3.5 Embed (1B and 8B) worth it?
- Nemotron 3.5 Embed is the open embedder to start with for document search, because it reads page images as well as text and costs nothing to try. We rate it 4.4 out of 5.
- How much does Nemotron 3.5 Embed (1B and 8B) cost?
- Free under OpenMDW-1.1. Both sizes are free to call on OpenRouter; Baseten, DeepInfra and FriendliAI also host them, and NVIDIA offers a NIM container.
- What are the downsides of Nemotron 3.5 Embed (1B and 8B)?
- All benchmarks are NVIDIA's own; linux and NVIDIA GPUs only; text evaluated in 34 languages, images in 18; queries must be text; image queries are not supported; version-sensitive: image input breaks on Transformers 5.17.0 to 5.19.0 with FlashAttention 2.
- Who is Nemotron 3.5 Embed (1B and 8B) best for?
- Teams building search or RAG over PDFs, slides and reports who want one open model for both text and page images.
- How much does Nemotron 3.5 Embed (1B and 8B) cost for a team or in production?
- Weights are free and OpenRouter currently charges nothing; self-hosted, NVIDIA measured the 1B NVFP4 at 128.8 documents a second on a GB200 and NIM at 6,031 query requests a second on an RTX PRO 6000.
- Is Nemotron 3.5 Embed (1B and 8B) safe for company data?
- OpenMDW-1.1 allows commercial use. Self-hosted, documents stay on your hardware. Training used public, commercially permissible datasets and synthetic data from models including Nemotron, Qwen, Gemma and gpt-oss.
- What are the best Nemotron 3.5 Embed (1B and 8B) alternatives?
- The closest alternatives we have reviewed are Qwen3-VL-Embedding, Nemotron ColEmbed V2, Cohere Embed 5, Harrier-OSS-v1 (270M, 0.6B, 27B) and Jina Embeddings v5.
Answers checked Nov 7, 2026.
Make Nemotron 3.5 Embed (1B and 8B)? Show this rating on your site
