Nvidia releases Nemotron 3.5 Super VL: Tops U.S. open-source model rankings, lags Xiaomi by nearly 20 points in benchmark scores

Beating AI News Flash: Nvidia has launched Nemotron 3.5 Super VL, releasing its model weights as open-source. The model has 120 billion total parameters, with only 12 billion activated per run, supporting a maximum 1 million-token context window. It combines text reasoning, visual understanding, and tool calling, allowing agents to directly read PDFs, analyze charts and videos, search the web, and complete tasks via tool calls—eliminating the need for repeated information transfer between text and vision models. The new model adopts a hybrid Mamba-Transformer architecture, improving inference efficiency through multi-token prediction and 4-bit quantization. In tests by Artificial Analysis, with a 100,000-token input, it achieved an output speed of roughly 630 tokens per second, about 5 times faster than Inkling Small, a model with comparable capability scores. Nvidia also noted that batch processing 10,000 SWE-bench Pro programming tasks using 4 B200 GPUs is up to 4 times faster than some comparable open-source models. Artificial Analysis gave Nemotron 3.5 Super VL a score of 26.5, ranking it first among U.S. open-source models, outperforming Nvidia’s earlier 550-billion-parameter Nemotron 3 Ultra (23 points) but still behind Xiaomi’s MiMo-V2.6-Pro (46 points). In Nvidia’s published SWE-bench Pro comparison, the new model’s pass rate stood at 51.78%, lower than Qwen3.8 Flash-Next’s 60.37%. Independent testing also exposed gaps between speed and reliability: Plotly evaluated the model on 43 data analysis questions, with the model answering 30 correctly, a median time of just 12 seconds per question, but missing all 3 data quality questions and 5 hidden pattern questions.