Mindbeam and NVIDIA Target Faster Open Source AI Performance
By refining NVIDIA’s Nemotron 3.5 Super VL model, New York-based startup Mindbeam has created a system capable of generating high-performance CUDA kernels. This collaboration addresses the growing engineering challenge of automating low-level code generation while maintaining strict numerical tolerances across diverse hardware architectures.
The newly developed Litespark Nemotron 3.5 Super VL model demonstrates significant gains in coding proficiency. In internal KernelBench evaluations, the model improved its success rate on individual operations from 72.7% to 83.9%. More notably, the model’s performance on complex, multi-level tasks surged from 22.8% to 82.1%, effectively surpassing the 57% accuracy mark set by GPT 5.6 Sol.
To complement this model, Mindbeam is deploying first-class support within its Litespark Inference Engine. When running on two NVIDIA H200 GPUs, the engine reaches throughput speeds of 947 tokens per second, with 962 tokens per second per user in warm-start benchmarks. According to Nii Osae, founder and CEO of Mindbeam, the initiative aims to make frontier-class AI capabilities more accessible by pairing custom-trained models with hardware-optimized serving infrastructure.
Comments (0)
No comments yet. Be the first!