Post from NVIDIA AI linking to x.com

As the need to optimize low level systems and kernel designs increases, modern frontier open source language models write CUDA routinely. The strongest engineering question is: can a model read a PyTorch module and write a raw CUDA kernel that computes the same forward pass within a fixed numerical tolerance while remaining architecture agnostic? To address this problem, Mindbeam in collaboration with @NVIDIAAI , is thrilled to launch a post-trained version of the frontier model Nemotron 3.5 Super VL 120B-A12B. The outcome is a powerful open-weights model Litespark Nemotron 3.5 Super VL that excels in CUDA code across various levels of KernelBench. On single operations in the Level 1 evaluation, the post-trained model passes 83.9% of problems, up from 72.7% before adaptation: a gain of 11.2% points from fine-tuning. Results are even stronger in the Level 2 evaluation where the model's performance improved from 22.8% to 82.1%, outperforming GPT 5.6 Sol's accuracy of 57%. This achievement marks a significant milestone for us, bolstering our vision of solving the hardest engineering problems in AI infrastructure. Learn more: https://mindbeam.ai/research/litespark-nemotron-3-5-super-vl