Mindbeam claims Litespark serves Nemotron 3.5 Super VL at 962 tokens/s and boosts CUDA accuracy

Mindbeam says it post-trained NVIDIA's Nemotron 3.5 Super VL 120B-A12B to write raw CUDA kernels from PyTorch modules, raising KernelBench Level 2 accuracy from 22.8% to 82.1%, above GPT 5.6 Sol's 57%.

It also says its Litespark Inference Engine serves the model at 962 tokens per second per user on two NVIDIA H200 GPUs, a 6.9x improvement over vLLM, with 947 tokens per second of output throughput in a warm-start benchmark. Mindbeam adds that the engine has tunable parameters for adjusting deployment performance. The company reports a Level 1 pass rate of 83.9%, up from 72.7% before adaptation. These figures are Mindbeam's own and have not been independently verified.

NVIDIA AI reposted Mindbeam's announcement on X, but the posts carry the same company claims and add no independent measurement.