NVIDIA
NVIDIA opens Nemotron 3.5 Super VL, a 120B model that reads video
NVIDIA published the weights for Nemotron 3.5 Super VL, a 120B model that reads text, images and video, under a license that allows commercial use.
NVIDIA has published the weights for Nemotron 3.5 Super VL, a model that reads text, images and video. The license allows commercial use. Open weights means anyone can download the trained model and run it on their own hardware.
The Model Activates 12 Billion of Its 120 Billion Parameters
Nemotron 3.5 Super VL has 120 billion total parameters, of which 12 billion are active for each token. A mixture-of-experts design splits a model into many specialist sections and sends each token through only a few. Most of the model sits idle on any one step. NVIDIA built this one as a hybrid of Mamba-2 layers, mixture-of-experts layers and attention layers, with a C-RADIO vision encoder.
The model accepts a context of up to 1 million tokens and outputs text. NVIDIA describes it as built for reasoning, coding, tool use, document understanding and video understanding. Reasoning can be turned off or on. The model also includes Multi-Token Prediction for speculative decoding.
NVIDIA Created Three Repositories on Hugging Face
NVIDIA created three repositories under its nvidia organization on Oct. 12, according to the model card on Hugging Face, last modified Oct. 22. They hold a full-precision BF16 version, a Base-BF16 version and an NVFP4 version.
The license is OpenMDW-1.1. The card says the model “is ready for commercial or non-commercial use.”
Full Precision Needs Four H100 GPUs or Two GB200 Chips
The card lists 4 H100 GPUs or 2 GB200 chips for the BF16 weights. It points to the NVFP4 release for efficient inference on a DGX Spark or data center GPUs. Its example serves the model with vLLM, an open-source inference server, at tensor parallelism of 4. That setting splits the model across four GPUs.
NVIDIA’s card reports 89.52 on GPQA-Diamond and 71.10 on SWE-bench Verified for the BF16 model. Those are NVIDIA’s own figures, as of the Oct. 22 update. The card’s table shows Qwen3.5-122B-A10B at 86.7 and 71.9 on the same two tests.
OpenRouter Lists It at $0.15 per Million Input Tokens
OpenRouter added the model on Oct. 15. Its model list on Oct. 25 shows $0.15 per million input tokens and $0.50 per million output tokens, with cached input reads at $0.02. The context window is 1,048,576 tokens, and a free variant is also listed. Prices get cut.
Our View
The card asks for four H100s at full precision. All 120 billion parameters have to sit in GPU memory, though only 12 billion work on each token. The active count cuts the compute per token and leaves the memory bill alone.
A company hosting the model can buy that capacity, use the NVFP4 version the card points to for efficient inference, or rent by the token. Developers testing the model can skip the hardware and pay OpenRouter’s per-token rate.
Occasional requests favor the per-token rate. Owned GPUs pay off only if steady traffic keeps them busy.