Amazon SageMaker adds NVIDIA Nemotron 3.5 Super VL model

Amazon tells that NVIDIA Nemotron 3.5 Super VL is available in Amazon SageMaker JumpStart. This tool deploys models in a small number of steps. The model reads images, video and text, and it writes text. NVIDIA made it for documents, screen tasks of agents that use computers, and video questions. The model has 120 billion parameters.

What you get

There are two variants. NVFP4 is a smaller version for G7 instances, and it is the default. BF16 has the full precision weights and operates on P5en instances. Each variant holds context of 1 million tokens at most.

The numbers

NVIDIA model cards show that NVFP4 matches BF16 on MMLongBench-Doc. It is in one point on the other tests. In a test by Amazon, BF16 gave 140.85 tokens each second for each query. With 10,000 input tokens, the first token took 12.62 seconds on ml.g7.24xlarge.

What is not known

The two variants operated on different hardware, so the test does not compare the two precisions. The endpoint accepts one image and one video for each query. It does not install LoRA adapters, which are small tuning files. New accounts start with a quota of zero for G7 instances.

Insight: NVIDIA gives the model away and sells the chips

In October, Nscale added the model, and NVIDIA tells that it operates on one Blackwell GPU. Cassava also added Nemotron models. Amazon runs it on G7 and P5en instances, which use NVIDIA GPUs. This reads as a plan to put open models on many hosts, so that each new host buys NVIDIA chips. Leaders must compare hosts by chip cost and not by model.