Hyperkernel generates SOTA megakernels across 11 domains, now on GitHub
MillionInt · x · 2026-11-06
Hyperstition unveiled Hyperkernel, a recursive megakernel generator claiming state-of-the-art performance across 11 domains — inference (DeepSeek, Kimi, GLM), pretraining (Qwen), hardware fault recovery, robotics (π₀-FAST, ACT), diffusion and real-time world models (Matrix-Game), FSD (OpenPilot), voice (Moshi), genomics (Evo 2), and protein folding (ESMFold) — all in a single kernel launch.
The team says fully optimized megakernels for robotics, genomics and voice were generated in under 50 minutes, roughly doubling efficiency for several models. Public implementations are on GitHub (MIT, Alpha), currently compiled for B300s, requiring Rust and NVIDIA driver R580+; users must supply their own checkpoints and tokenizers.
More from Infra
- J.B. Huang's new video explains how LLM serving actually works, from fast to cheap — jbhuang0604 · 2026-11-06
- US BIS extends suspension of affiliates end-user rule through January 10, 2027 — pstAsiatech · 2026-11-06
- Meta's Axia Learns Scaling Laws for Ranking Models from a Few Trial Runs — _reachsumit · 2026-11-06
- US trade deficit ex-data centers is just 0.7% of GDP, lowest since 2000 — soumitrashukla9 · 2026-11-06
- Modal's GPU Glossary praised as an excellent free reference — eigenron · 2026-11-06
- Underdog Woof: 31.7MB open-weight TTS model claims 1.42% WER, runs fully in-browser — _sathvikr · 2026-11-06