What can the five NVIDIA models do?

They cover embedding, reranking and parsing: Embed-8B and Embed-1B turn text and images into vectors, Rerank-8B ranks results by relevance, and Parse turns PDFs into AI-ready data.

How does the small model compare with the large one?

Nemotron-3.5-Embed-1B-BF16 is a scaled-down version of the 8B model, and NVIDIA says it performs close to Qwen3-VL-Embedding-8B. There is also an NVFP4 quantized version of the 1B model.

Do I have to pay to use them?

No. All five models are released free as open models.

Large Language ModelsOpen-Source AIGIGAZINE AI3h agoOct 22, 2026, 16:00 JST

NVIDIA releases five Nemotron-3.5 models free, led by Embed-8B

Send on LINE

3 Key Points

What happened
NVIDIA added five models to Nemotron Retriever and released them free as open models. Nemotron-3.5-Embed-8B-BF16 handles 34 languages including Japanese plus image and text input, and beat Qwen3-VL-Embedding-8B and EmbeddingGemma-2 on embedding benchmarks.
Why it matters
The benchmarks position these models as fast and accurate alternatives to competing embedding models, with the smaller variants showing performance close to larger rivals. That may make self-hosted retrieval setups cheaper to run, though NVIDIA's own presentation of the results is the main evidence offered.

WHO IT HITSEnterprise teams building self-hosted search and document-processing pipelines get a free, open set of models they can download and run themselves, rather than paying a vendor API.

Not sure about something? Ask the AI

Questions and answers are published on this page.

Summaries like this, in your inbox every morning.

Get news like this every morning, free

Context & Analysis

NVIDIA is extending its Nemotron Retriever line, which is already aimed at retrieval-augmented generation, with five new models spanning embedding, reranking and PDF parsing. The company put its benchmark comparisons front and center: the 8B embedding model is shown above Qwen3-VL-Embedding-8B and EmbeddingGemma-2, and the 1B models are shown tracking closely to Qwen3-VL-Embedding-8B despite their smaller size.

NVIDIA also highlights the trade-off between speed, accuracy and cost. Graphs in the post plot processing speed against accuracy and per-query cost against accuracy, with the three embedding models shown as fast and accurate relative to competitors. It says pairing Nemotron-3.5-Embed-8B-BF16 with Nemotron-3.5-Rerank-8B-BF16 cuts cost while improving accuracy.

On the parsing side, Nemotron-3.5-Parse is presented as both accurate and fast, with its pages-per-second rate standing out in NVIDIA's comparison. The release is free and open, so organizations can evaluate these claims on their own document workloads rather than take the charts at face value.

FAQ

What can the five NVIDIA models do?: They cover embedding, reranking and parsing: Embed-8B and Embed-1B turn text and images into vectors, Rerank-8B ranks results by relevance, and Parse turns PDFs into AI-ready data.
How does the small model compare with the large one?: Nemotron-3.5-Embed-1B-BF16 is a scaled-down version of the 8B model, and NVIDIA says it performs close to Qwen3-VL-Embedding-8B. There is also an NVFP4 quantized version of the 1B model.
Do I have to pay to use them?: No. All five models are released free as open models.

GIGAZINE AIRead Original Article

Send on LINE

AI news that matters for your work, delivered every morning.

Pick your industry and the AI tools you use, and get news related to your work every day.

Get it free

Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →

Ask AI

Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.