Back home

Hyperkernel

x+github 0.88
Visit source

Intro

Hyperkernel is a recursive GPU megakernel generator open-sourced by lab Hyperstition (formerly Social Physics Lab) on 2026-11-05/06. GitHub: https://github.com/Hyperstitioncc/Hyperkernel ; site: https://hyperstition.cc/ .

They claim ~60–100 lines of description can yield a model-specific megakernel in under an hour, fusing a whole inference/training path into one kernel launch, across 11 domains including DeepSeek / Kimi / GLM inference, Qwen pretraining, robotics (π₀-FAST, ACT), world models, OpenPilot, Moshi voice, Evo 2 genomics and ESMFold.

Versus NVIDIA's published vLLM numbers (self-reported):

  • DeepSeek V4.1 Flash: +36.7% per-user generation speed at concurrency 1; +90.6% at concurrency 4 vs the highest published interactivity figure;
  • ~3.2× FPS on real-time world models, among other claims.

Repo created 2026-11-05, currently Alpha; primary target NVIDIA B300 (with stated paths to B200 / AMD / TPU). Full generator stack and Kimi/GLM serving planned over coming weeks.

Explain

Why it matters

Megakernels collapse chains of short kernels and launch overhead into one persistent launch — critical for low-batch / low-latency work (voice, robotics, agent loops). Hyperkernel claims to automate that, so labs without a dedicated CUDA team can chase rack-level efficiency.

Credibility caveats

Numbers are self-reported against "NVIDIA published" baselines; independent reproduction is thin so far. Repo is Alpha; stars and docs are moving fast.

Versus peers

RightNow-AI's AutoMegaKernel and similar agent-driven synthesizers sit in the same lane; Hyperkernel's pitch is a recursive generator plus multi-domain megakernels shipping on day one.

Usage

Who it's for

  • Systems engineers optimizing LLM inference or low-latency robotics/voice serving;
  • Researchers tracking GPU megakernels / single-launch fusion;
  • Teams with B300/B200 access willing to run Alpha code.

Getting started

  1. Read https://github.com/Hyperstitioncc/Hyperkernel README and hardware needs (Rust, NVIDIA R580+, etc.);
  2. Bring your own model checkpoint and tokenizer;
  3. Build/load the shipped megakernels per the repo and benchmark yourself against the team's claims;
  4. Watch for the full generator stack and Kimi/GLM serving releases.

Advertisement