NVIDIA releases Nemotron 3.5 Super VL, a 120B open model

Watch

Lab claim

On October 15, 2026, NVIDIA released Nemotron 3.5 Super VL, an open model with 120 billion parameters, and a 24 GB edge model it calls Bolt 74B-A9B.

How to read this claim

Measurement — The post does not name the Hugging Face repository. The 5-times comparison does not name the other model in the prose.
Bill — List input and output per million tokens on the vendor page — a loop pays both again.
Door — Status on this card: Shipped.

Keep the habit · field card with prices and worked checks

What changed

NVIDIA's open agent lineup added a 120 billion parameter multimodal model and a 24 GB edge build on October 15. The 4 GB build is announced for October 23.

What it introduces

An open 120 billion parameter agent model and a 24 GB quantized edge model. The 4 GB model is not available yet.

Who benefits Anyone placing an open agent model on NVIDIA hardware, and anyone who needs the 24 GB build rather than the 120 billion parameter model.

Try this

Read the 24 GB Bolt note before you assume Super VL fits a desktop. Treat the 5-times throughput line as NVIDIA's. The post says the weights are on Hugging Face and does not print the repository name on the page we read.

What we do not know

The post does not name the Hugging Face repository. The 5-times comparison does not name the other model in the prose.

Evidence

Primary NVIDIA. Reported by NVIDIA. The 5-times throughput comparison is NVIDIA's reading of an Artificial Analysis chart.

Independent The throughput comparison is NVIDIA's, drawn from an Artificial Analysis chart. No independent rerun is cited on the post.

Sources

NVIDIA Primary · October 15, 2026

Still unclear? · Start here · ranks, prices, unknown The post does not name the Hugging Face repository. The 5-times comparison does not name the other model in the prose.

The record

What happened

On October 15, 2026, NVIDIA released Nemotron 3.5 Super VL, which the post describes as an open 120 billion parameter model, and Nemotron 3.5 Bolt 74B-A9B in a 24 GB build that uses native 2-bit quantization. A third model, Bolt E4B at 4 GB, is dated October 23 and is not available on this page. NVIDIA says Super VL was trained on about 26 trillion tokens and then post-trained with roughly 3.8 million samples. NVIDIA also says Super VL reaches up to 5 times the throughput of other models in its class on an Artificial Analysis chart.

Why it matters

Beam and Muse Spark Open are the open agent files from October 15 and 16. Super VL is a third open file, from NVIDIA, with a named edge build that the post says fits in 24 GB. The 5-times figure is NVIDIA's comparison, not a named opponent in the sentence, so it is not yet a reason to retire another model. The 4 GB Bolt is a date, not a download.

NVIDIA says the model ships with open weights, open datasets, and training recipes, and that it can be served with NIM and Dynamo. The 8-times user and 85 percent latency lines on Dynamo are NVIDIA's.