The claim worth reading twice

Most model announcements sell accuracy. This one sells a running cost. NVIDIA's experiment is better designed than the average vendor chart: an agent searched a corpus, inspected what came back and refined its queries, while researchers swapped only the retrieval layer underneath. The language model, the tools, the task set and the corpus stayed fixed, and the meter counted total agent inference tokens — input and output across all turns. NVIDIA names the resulting metric "agentic token cost per query", and on that metric adding the reranker reduced the cost by approximately 14% alongside the 8B embedder and 13% alongside the 1B.

The proposed mechanism is plausible: an agent that receives the right page early could stop paying for extra search and reasoning rounds. But that is an outcome observed inside this harness, not proof of a universal mechanism, and it should not be converted into money. Tokens are not priced uniformly — input, output, cached and tool tokens all price differently — so a 14% token reduction does not translate into a 14% reduction in anyone's bill. Buyers should reproduce the full-agent result on their own queries before treating it as a budget assumption.

The same section of the announcement makes a second, distinct comparison that is easy to conflate with the leaderboard numbers below. With Embed 8B as the retriever inside the agent evaluation, Rerank 8B delivered approximately 0.9 points higher nDCG@10 than Qwen3-VL-Reranker-8B at a similar agentic token cost. A saving measured against one baseline, in one harness, tells you the direction of the effect. It does not tell you the size of the effect in your harness, against the reranker you already run.

A leaderboard launched by its winner

The retrieval scores need the same reading. Embed 8B's 61.17 nDCG@10 on ViDoRe v3.1 is a substantial reported difference over the 52.23 of NVIDIA's previous 1B visual embedder, the Llama Nemotron Embed VL 1B, and its first place among dense retrieval models is NVIDIA's own scoped chart claim, dated 20 October 2026. The distilled 1B model retaining 95.4% of the 8B score (58.38) is, on NVIDIA's reporting, the most useful engineering fact in the release: multimodal retrieval at a size that can be indexed cheaply.

The reranking crown is softer. NVIDIA is launching the ViDoRe v3.1 reranking leaderboard together with the MTEB team, and reports its own model leading it at launch: 60.49 nDCG@10 on combined image-and-text inputs against 59.17 for Qwen3-VL-Reranker-8B — a 1.32-point difference, and a separate result from the 0.9-point agent-evaluation comparison above. Vendor participation does not invalidate a benchmark. What will determine confidence in this one is the ordinary machinery: public datasets, reproducible scoring and submissions from outside the vendor.

The partner results published inside the announcement sit in the same category. They are partner-reported results that NVIDIA chose to publish, not independently reproduced public benchmarks, and several of the partners also distribute or support the models. Turbopuffer reports about 5% more relevant results at Recall@20 than the next-best model on an enterprise PDF benchmark, and now sells the models as a native option. Vectorize ranks Embed 8B first on two long-conversation memory tests, LoCoMo and BEAM 100K. Perplexity reports Rerank 8B first on a private 72-task evaluation, which no reader can inspect. A partner evaluation published inside the vendor's announcement is not independent verification. That does not make the results false; it makes them claims awaiting an outside rerun.

Parse is the quiet release

The least promoted model may travel furthest, and it is a concrete offer rather than a vision statement for a simple reason: Parse has been released, and anyone can test it. At 631 million parameters it reads a document page and returns its structure — reading order preserved, tables as HTML or Markdown tables, charts as structured blocks. Its modes trade accuracy for speed in plain numbers, all on one H100: 72.74 on ParseBench at 16.1 pages per second in a single pass, and 75.20 at 10.5 pages per second in triple pass. NVIDIA describes it as first among models under 1 billion parameters on ParseBench. Those speeds are vendor-reported under specified serving conditions, not guaranteed production throughput — but document ingestion is the unglamorous bottleneck in most enterprise retrieval projects, and a sub-billion-parameter open model that does it at ten-plus pages per second per GPU can be measured on your own documents this week.

Two smaller claims deserve the discount instead. The fine-tuning recipe's headline gain — 18% relative nDCG@10 for the 1B model, 3% for the 8B — was measured on held-out queries that were themselves synthetic, generated by Qwen3.8-Flash-Next over chart and infographic images; held-out is not the same as real-world, and synthetic queries flatter a recipe tuned on synthetic data. And the NVFP4 variant's 27.1% throughput advantage over its BF16 sibling was measured on a GB200, NVIDIA's own test configuration, on hardware most teams reading this do not have. Neither claim is hidden. Both are printed next to their conditions, which is more than most releases manage.

One licence note, because it decides what "open" means here. Every repository in the suite states OpenMDW-1.1 — checked on 21 October — and the Embed 8B card describes the licence as permitting commercial and non-commercial use. The underlying Ministral-3-8B component, however, carries a separate Apache 2.0 licence. Read the licence attached to the repository you actually download: commercial permission under OpenMDW-1.1 is not the same as unrestricted use.

Sources

  1. [1] NVIDIA, "NVIDIA Nemotron 3.5 Embed, Rerank, and Parse: Advancing Agents with World-Class Retrieval Models", Hugging Face blog, 21 October 2026 — the release, the agent token-cost experiment, all benchmark and throughput figures, partner evaluations, and availabilityRead source
  2. [2] NVIDIA, Nemotron-3.5-Embed-8B-BF16 model card, Hugging Face — OpenMDW-1.1 licence, commercial-use statement, release date 10/20/2026, 8.38B parameters, 32,768-token contextRead source
  3. [3] NVIDIA, Nemotron-3.5-Embed-1B-BF16 model card, Hugging Face — licence field states OpenMDW-1.1 (checked 21 October 2026)Read source
  4. [4] NVIDIA, Nemotron-3.5-Rerank-8B-BF16 model card, Hugging Face — licence field states OpenMDW-1.1 (checked 21 October 2026)Read source
  5. [5] NVIDIA, Nemotron-3.5-Parse model card, Hugging Face — licence field states OpenMDW-1.1 (checked 21 October 2026)Read source
  6. [6] Unite.AI, "NVIDIA Debuts Open Nemotron 3.5 Models for Embedding, Reranking, Parsing", 21 October 2026 — independent write-up of the same announcement, used as a cross-checkRead source