Hyperstition open-sources Hyperkernel megakernels it says beat NVIDIA's published DeepSeek V4.1 Flash speeds
The MIT-licensed alpha repo ships generated megakernels for B300 GPUs; the company's own benchmarks claim up to 90.6% higher per-user speed than NVIDIA's published vLLM results, unverified, and the generator itself is promised in a couple of weeks.
Why it matters
A megakernel fuses an entire model's forward pass into one GPU launch, cutting launch and synchronisation overhead that dominates batch-one latency. Hyperstition says its Hyperkernel generator produced kernels for DeepSeek V4.1 Flash, π0-FAST, ACT, Moshi, Evo 2 and others in under an hour for some models, and the repo README confirms the released builds target four B300s for LLMs and need Rust and an NVIDIA R580+ driver. The speed figures come from the company's own thread; the full stack and Kimi and GLM serving code are promised later, so independent reproduction is the thing to watch.
Key facts
- Repo Hyperstitioncc/Hyperkernel, MIT licence, alpha; last public commit 2026-11-05T22:44Z
- Released megakernels target NVIDIA B300s (4 GPUs for LLMs); needs Rust and driver R580+
- Claimed 36.7% higher per-user speed at concurrency 1 and 90.6% at concurrency 4 on DeepSeek V4.1 Flash vs NVIDIA's published vLLM results
- Claimed π0-FAST 94 ms vs 243 ms and ACT 0.81 ms vs 2.21 ms per frame
- Full Hyperkernel stack and Kimi/GLM serving stack promised in a couple of weeks
- Benchmark figures are the company's own, not independently verified