Nvidia Releases Nemotron 3.5 Super VL: No. 1 Open Source in the U.S., but Benchmark Score Trails Xiaomi by Nearly 20 Points

Beating AI News Flash: NVIDIA releases Nemotron 3.5 Super VL, simultaneously open-sourcing the model weights. It has 120 billion total parameters, activating only 12 billion per pass, and supports up to 1 million Token context.

The model combines text reasoning, visual understanding, and tool calling. Agents can directly read PDFs, analyze charts and videos, then search the web and call tools to complete tasks, without repeatedly passing information between text models and vision models.

The new model continues the hybrid Mamba and Transformer architecture, and improves inference efficiency through multi-Token prediction and 4-bit quantization. Artificial Analysis measured an output speed of approximately 630 Tokens per second at 100,000 Token input, about 5 times that of Inkling Small, which has a similar capability score.

NVIDIA also claims that using 4 B200 GPUs to batch process 10,000 SWE-bench Pro coding tasks is up to 4 times faster than some comparable open-source models.

Artificial Analysis gives it a score of 26.5, ranking first among U.S. open-source models, surpassing NVIDIA's previous 550 billion parameter Nemotron 3 Ultra (23 points), but still trailing Xiaomi MiMo-V2.6-Pro's 46 points. In the SWE-bench Pro comparison published by NVIDIA, the new model's pass rate is 51.78%, lower than Qwen3.8 Flash-Next's 60.37%.

Independent testing also exposed the gap between speed and reliability. Plotly tested it with 43 data analysis questions; the new model answered 30 correctly, with a median time of only 12 seconds per question, but missed all 3 data quality questions and 5 hidden pattern questions.