Mindbeam said Thursday it post-trained NVIDIA’s Nemotron 3.5 Super VL for CUDA kernel generation and added support for NVIDIA’s model to its Litespark Inference Engine. It reported higher GPU coding test scores for the adapted version.
NVIDIA said it released the model as open on Oct. 15. Its technical blog also lists Mindbeam among organizations that post-trained Nemotron 3.5 Super VL.
Mindbeam reported that its adapted model passed 83.9% of KernelBench Level 1 tasks, compared with 72.7% before adaptation, a gain of 11.2 percentage points. The company said its Level 2 result rose from 22.8% to 82.1%. It also said the Level 2 score exceeded the 57% it reported for “GPT 5.6 Sol,” but gave no test conditions for that comparison.
In a warm-start test using two NVIDIA H200 GPUs, Mindbeam said its system produced 962 tokens per second per user and output throughput of 947 tokens per second.
The release did not identify the KernelBench version, test hardware and software, scoring method or number of runs. For the H200 benchmark, Mindbeam did not specify the workload, model precision, batch or concurrency settings, or input and output lengths, making it difficult to compare those rates with other inference tests.
KernelBench tests whether a model can produce GPU code that works correctly and efficiently for tasks expressed in PyTorch. Its first tier covers 100 individual operator problems; the second covers 100 problems that combine operations through simple fusion. The project documentation describes the generated code as CUDA or domain-specific language (DSL) kernels.
NVIDIA describes Nemotron 3.5 Super VL as an open model with 120 billion parameters, 12 billion of them active per token. It handles text, images, documents and video.
Mindbeam framed the collaboration as a way to improve speed and efficiency in open-source AI development and deployment. The company did not identify a customer or quantify possible savings.
