When was Nemotron 3.5 Super released?

NVIDIA lists 15 October 2026 as the Hugging Face release date for Nemotron 3.5 Super VL. The technical report followed on arXiv on 28 October 2026.

How large is Nemotron 3.5 Super?

120 billion total parameters, of which 12 billion are active per token, with a context of up to 1M tokens.

Can Nemotron 3.5 Super be used commercially?

Yes. The model card says it is ready for commercial and non-commercial use under the OpenMDW License Agreement 1.1.

Nemotron 3.5 Super news

Open-weight vision-language model1 story

Nemotron 3.5 Super VL is an open-weight vision-language model from NVIDIA, published on Hugging Face on 15 October 2026.

An NVIDIA DGX Spark desktop computer with a gold metal-foam front panel on a wooden table. — NVIDIA details Nemotron 3.5 Super and Lightning, open hybrid Mamba MoE models for agents

Open Source AI31 Oct 2026, 14:05 UTC

NVIDIA's report on Nemotron 3.5 Super (120B, 12B active) and Lightning (30B, 3B active) explains how the open hybrid Mamba-Transformer models were trained.

Summary

NVIDIA published the technical report for Nemotron 3.5 Super and Nemotron 3.5 Lightning on arXiv on 28 October 2026. Super is a 120B-parameter multimodal model with 12B active parameters; Lightning is a 30B text model with 3B active. Both are open mixture-of-experts models that mix Mamba and Transformer layers. NVIDIA reports 51.78% on SWE-bench Pro for Super and releases base, BF16 and NVFP4 checkpoints, datasets and recipes under OpenMDW 1.1.

About Nemotron 3.5 Super

Nemotron 3.5 Super VL is an open-weight vision-language model from NVIDIA, published on Hugging Face on 15 October 2026. It is a hybrid Mamba-Transformer mixture-of-experts model with 120B total and 12B active parameters, a context of up to 1M tokens and text, image and video input, built for reasoning, coding and agent work.

Reviewed 31 Oct 2026 · by NVIDIA

Facts at a glance

Developer: NVIDIA
Released: 15 October 2026 on Hugging Face source
Parameters: 120B total, 12B active (mixture of experts, 512 routed experts per MoE layer) source
Architecture: Interleaved Mamba-2, mixture-of-experts and attention layers with a C-RADIO vision encoder and multi-token prediction source
Context window: Up to 1M tokens source
Inputs and outputs: Text, images and video in; text out
Checkpoints: Base, BF16 post-trained and NVFP4 quantized, plus a LiLiCorr speculative drafter source
SWE-bench Pro: 51.78% (NVIDIA figure) source
Serving hardware in the model card: BF16 on 4x H100 or 2x GB200; NVFP4 for DGX Spark and data center GPUs source
License: OpenMDW License Agreement 1.1

More about Nemotron 3.5 Super (show)

What Nemotron 3.5 Super is used for

NVIDIA positions Super as the larger of the two Nemotron 3.5 general models, for reasoning, software engineering, tool use, document and video understanding. The technical report, submitted to arXiv on 28 October 2026, describes 1 trillion tokens of distillation from Nemotron 3 Ultra Base, Indic continued pretraining, 67B tokens of software-engineering mid-training and a post-training pipeline with reinforcement learning and multi-teacher distillation. NVIDIA also compressed Super VL into the smaller Nemotron 3.5 Bolt.

Questions people ask

When was Nemotron 3.5 Super released?: NVIDIA lists 15 October 2026 as the Hugging Face release date for Nemotron 3.5 Super VL. The technical report followed on arXiv on 28 October 2026.
How large is Nemotron 3.5 Super?: 120 billion total parameters, of which 12 billion are active per token, with a context of up to 1M tokens.
Can Nemotron 3.5 Super be used commercially?: Yes. The model card says it is ready for commercial and non-commercial use under the OpenMDW License Agreement 1.1.

Nemotron 3.5 Super news

Every story about Nemotron 3.5 Super

NVIDIA details Nemotron 3.5 Super and Lightning, open hybrid Mamba MoE models for agents