New York – October 15, 2026 -- Mindbeam raised the Level 2 KernelBench score of NVIDIA's Nemotron 3.5 Super VL from 22.8% to 82.1% through post-training, ahead of the 57% accuracy the company reports for GPT 5.6 Sol.
The New York-based AI infrastructure company announced a collaboration with NVIDIA covering two layers of the open source AI stack: a post-trained model that writes CUDA kernels, and an inference engine that serves NVIDIA's latest open models.
Litespark Nemotron 3.5 Super VL gains 11.2 points on Level 1 tasks
Mindbeam post-trained NVIDIA's Nemotron 3.5 Super VL (120B-A12B) to produce Litespark Nemotron 3.5 Super VL, an open-weights model for CUDA code generation. The task is to take a PyTorch module and generate a raw CUDA kernel that reproduces the forward pass within tight numerical tolerances while staying architecture-agnostic.
On KernelBench Level 1, which tests individual operations, the model passes 83.9% of problems, up from 72.7% before adaptation, a gain of 11.2 percentage points. On Level 2, performance rose from 22.8% to 82.1%.
Litespark Inference Engine reaches 962 tokens per second per user on two H200 GPUs
Mindbeam also launched day-0 support for Nemotron 3.5 Super VL in its Litespark Inference Engine. On two NVIDIA H200 GPUs, the engine delivers 962 tokens per second per user and 947 tokens per second of output throughput in the company's warm-start benchmark.
The engine exposes tunable parameters so users can optimize for throughput, latency or hardware utilization across production deployments of open-source models.
Mindbeam CEO commits to building on NVIDIA's open models
Founder and CEO Nii Osae said the company is proud to build alongside NVIDIA's open models and plans to show further developments. Mindbeam frames the release as a step toward faster, more cost-effective and more accessible frontier-class open source AI.