Announcing Mindbeam’s Collaboration with NVIDIA: Bringing Greater Speed and Efficiency to Open Source AI

We are excited to announce a collaboration with NVIDIA, spanning two critical layers of the open source AI stack: a post-trained frontier model that generates high-performance CUDA kernels and an inference engine that runs NVIDIA's latest open models with exceptional speed and efficiency.

Litespark Nemotron 3.5 Super VL

As demand grows for optimized low-level systems and custom kernels, frontier open-source models are becoming increasingly capable of writing CUDA. But a key engineering question remains: Can a model take a PyTorch module and generate a raw CUDA kernel that computes the same forward pass within tight numerical tolerances while remaining fully architecture-agnostic?

To answer this question, we post-trained NVIDIA’s frontier model Nemotron 3.5 Super VL (120B-A12B). The result is Litespark Nemotron 3.5 Super VL, a powerful open-weights model that excels at CUDA code generation across the different levels of KernelBench.

In the Level 1 evaluation, which tests individual operations, the post-trained model passes 83.9% of problems, up from 72.7% before adaptation. That’s an improvement of 11.2 percentage points through fine-tuning. Results are even stronger in the Level 2 evaluation where the model’s performance improved from 22.8% to 82.1%, outperforming GPT 5.6 Sol’s accuracy of 57%.

Day-0 Launch Support for Nemotron 3.5 Super VL

A model is only as useful as the engine that serves it. As state-of-the-art open models continue to multiply, so does the demand for high-performance inference. Every token per second matters and every millisecond of latency counts.

We are equally excited to announce first-class support for NVIDIA’s Nemotron 3.5 Super VL in the Litespark Inference Engine. The results speak for themselves: running on just two NVIDIA H200 GPUs, Litespark achieves 962 tokens per second per user and 947 tokens per second in output throughput in our warm-start benchmark.

The Litespark Inference Engine also supports a wide range of tunable parameters, allowing users to optimize performance for their specific workloads. Whether optimizing for throughput, latency or hardware utilization, Litespark provides the flexibility to support a diverse range of production scenarios for open-source model deployments.

We have developed a model that can generate performance-critical systems code, paired with an engine designed to get the most out of modern accelerators.

This is the direction open source AI is heading: making frontier-class capabilities faster, more cost effective, and more accessible to everyone.

We are proud to be building this future alongside NVIDIA's open models, and we can't wait to show you what comes next.