We’ve collaborated with NVIDIA to bring day-zero support for NVIDIA Nemotron 3.5 Super VL (opens in a new tab), their latest vision-language model, to the River API.
As part of our evaluation and post-training work, we used the text-to-SQL recipe from our recent post to train Super VL to 92% on Arcwise-Plat-SQL, outperforming every untrained proprietary model we evaluated.
Even before training, Super VL demonstrated competitive performance against models many times its size, highlighting the potential of NVIDIA’s latest vision-language model.
Big Performance, Small Footprint
Out of the box, Super VL scores 80.7% on Arcwise-Plat-SQL, closely matching Kimi K2.6 despite having roughly one-eighth as many total parameters. It also comes within about two points of much larger models like GLM-5.3 Flash and DeepSeek V4.1 Flash.
At 120B total parameters, Super VL is the smallest mixture-of-experts model we’ve tested to clear 80% out of the box. It activates just 12B parameters per token, compared with Kimi K2.6’s 32B. Its smaller footprint makes it more practical to self-host, while its low active parameter count means less compute per generated token and greater potential inference throughput.
Out-of-the-Box Accuracy by Parameters
Mixture-of-experts models on Arcwise-Plat-SQL (498 questions), 8,192 tokens per turn
Results After Training
The chart compares each model’s accuracy before and after training with our recipe (opens in a new tab). Proprietary models are included as untrained baselines, evaluated using the same grading methodology.
Getting Started
Nemotron 3.5 Super VL is now available on the River API. Explore the docs to get started, check out NVIDIA’s model card (opens in a new tab) for technical details, or contact us to learn more.