Nemotron 3.5 Super VL Outperforms Peers in Retro Game Rules Benchmark
A user-submitted benchmark evaluation reveals that the Nemotron 3.5 Super VL model successfully recalls obscure retro video game rules from Anno 1602 without web search, outperforming several other local models that hallucinated details. This informal evaluation highlights the practical capabilities of open-weight vision-language models in handling niche historical knowledge without relying on external internet searches. It provides local AI enthusiasts with insights into how modern models compare when managing highly specific informational constraints. Running with NVFP4 quantization on dual RTX 5060 Ti 16GB GPUs, the Nemotron 3.5 Super VL achieved a generation speed of 8-10 tokens per second and processing speeds of 150-170 tokens per second while making only minor mistakes compared to major hallucinations from competing models.
## BACKGROUND
Vision-language models are multimodal AI systems capable of interpreting and generating information from both images and text simultaneously. Nemotron is a family of AI models developed by NVIDIA that includes multimodal architectures designed for advanced reasoning and information retrieval tasks.