Hyperstition open-sources Hyperkernel for faster GPU inference

A team called Hyperstition released an open-source project on GitHub named Hyperkernel, which automatically generates optimized megakernels, essentially large, highly tuned chunks of GPU code, for running AI models. The project claims substantial speed improvements compared to NVIDIA's own vLLM inference software when running large models similar to DeepSeek, based on the team's own benchmarks. It also reports faster frame rates for real-time world-model applications in its testing. Megakernel generation is a technique aimed at squeezing more performance out of existing GPU hardware by combining operations that would normally run as separate steps into a single, more efficient unit of execution. If the claimed speedups hold up under independent testing, tools like this could meaningfully lower the cost of running large AI models in production, since inference efficiency has become one of the biggest cost drivers for companies deploying these models at scale.

Why it matters

An open-source tool claiming major inference speedups could lower real-world costs of running large AI models.

Source: Hyperstition