NVIDIA Nemotron 3.5 Retriever collection: Embed 8B/1B, Rerank 8B, Parse 631M
NVIDIA's Oct 21 collection drop for agentic retrieval: Nemotron-3.5-Embed in 8B BF16, 1B BF16 and 1B NVFP4, Nemotron-3.5-Rerank-8B cross-encoder, and Nemotron-3.5-Parse (631M) for document image extraction. Open weights, data and training recipes on Hugging Face; commercial use under OpenMDW-1.1.
Why it matters
First retrieval stack built around agentic token economics: NVIDIA reports the reranker cuts an agent's total inference tokens per task by ~14% alongside the 8B embedder, and the 8B embedder tops ViDoRe v3.1 dense retrieval (61.17 nDCG@10) and MIRACL Vision (72.6). Parse scores 75.20 on ParseBench at 10.5 pages/sec on one H100 — first among sub-1B models.
What you could build with it
A startup building a document-heavy RAG product could chain Nemotron-3.5-Parse (scan-to-structured-Markdown) with the Embed + Rerank models to index contracts or research PDFs once, then measure whether the reranker's token savings materialize on real customer queries before committing to the stack.
Does it hold up?
Too early to judge — released Oct 20-21; the only evidence so far is NVIDIA's own benchmarks and partner evaluations (Elastic, Vectorize, Perplexity, Turbopuffer). No independent community bake-offs exist yet.
Built with NVIDIA Nemotron 3.5 Retriever collection
- NVIDIA Releases Nemotron 3.5 Retriever Models (Daily AI Brief)dailyaibrief · Independent rundown of the Oct 20-21 collection: model specs, Ministral-3-8B base, 32K context, OpenMDW-1.1 commercial licensing, and the fine-tuning recipes.
- NVIDIA Releases Nemotron 3.5 Parse for Document Extraction (Daily AI Brief)dailyaibrief · Writeup of the Oct 17 Parse release: C-RADIOv4-SO400M vision encoder + 10-layer mBART decoder, 75.20 ParseBench, 83.73 on IndicVisionBench OCR across 10 languages.
- Nemotron-3.5-Rerank-8B-BF16 in The Deploy Logdeployedbyai · Deployment record of the 8.38B cross-encoder reranker on Hugging Face with NVIDIA as publisher source, logged Oct 15.
Learn more
NVIDIA's technical deep dive on Hugging Face →
First spotted on huggingface: source.
More AI models
OpenAI publishes 722 AI-generated math manuscripts from an unreleased modelOpenAI released 722 mathematical manuscripts in 372 result families to GitHub under Apache-2.0, produced by an…models · JEV 0.74Tencent Hunyuan Hy3: open 295B MoE with 21B active under Apache 2.0Tencent open-sourced Hy3, a 295B-parameter mixture-of-experts model with only 21B parameters active per token, under…models · JEV 0.73NVIDIA Isaac Lab 3.0 GA + Isaac OS open runtime + Agile ONE SN humanoidMajor Isaac platform expansion at GTC Berlin Oct 21: Isaac Lab 3.0 is generally available as the open-source robot…models · JEV 0.73StepFun Step-5-Preview — open BF16 weights releasedStepFun published the full BF16 weights for its flagship Step-5-Preview agentic model (600B-parameter sparse MoE, 27B…models · JEV 0.73Kimi K3.1 Pro: Moonshot's 2.8T open-weight flagship officially launchesMoonshot AI officially launched Kimi K3.1 Pro on Oct 23, a mixture-of-experts model with 2.8 trillion total / 104…models · JEV 0.72ElevenLabs Eleven v4 + v4 Turbo: new emotive TTS architecture with a 100ms real-time variantElevenLabs released Eleven v4 and Eleven v4 Turbo on September 28: a new text-to-speech architecture with inline…models · JEV 0.72
Get the week's best AI launches, plus 3 ideas worth building
One email every Saturday. Ranked by traction, not hype. Free.