Post from NVIDIA AI linking to x.com

Today, we are thrilled to announce that Mindbeam has collaborated with @NVIDIAAI to optimise the speed and efficiency of open source models. We post-trained NVIDIA’s Nemotron 3.5 Super VL to translate PyTorch modules directly into raw CUDA kernels, moving KernelBench Level 2 accuracy from 22.8% to 82.1% and outperforming GPT 5.6 Sol's 57%.  To ensure the serving layer keeps pace, the Litespark Inference Engine delivers Day-0 support, with 962 tokens per second per user on just two NVIDIA H200 GPUs, delivering a 6.9x improvement over vLLM.  Generating performance-critical systems code is only half the battle; running it efficiently in production is the other. We are proud to be building both sides of that equation alongside NVIDIA. Learn more: https://mindbeam.ai/research/mindbeam-nvidia-nemotron-collaboration