July 19, 2026

Kimi K3 writes H100 CUDA kernel 14.82x faster than PyTorch

Kimi K3 has achieved a significant milestone by writing a CUDA kernel that is 14.82 times faster than optimized PyTorch, nearly matching the performance of Claude Opus 4.8. This accomplishment was evaluated using KernelBench, a platform that assesses AI models on their ability to generate custom GPU implementations that are correct and outperform reference implementations on actual hardware. Recent evaluations highlight Kimi K3’s competitiveness in advanced coding tasks, particularly those requiring low-level optimizations.

Kimi K3: Kimi K3 is Moonshot AI’s latest flagship large language model, released in mid-July 2026 as the company’s most capable offering to date with a focus on frontier-level coding, reasoning, and knowledge work. It features native vision capabilities and architectural innovations like Kimi Delta Attention for handling long sequences efficiently. In this news, Kimi K3 showcases its code generation strengths by producing a highly optimized H100 CUDA kernel that significantly outperforms PyTorch on KernelBench hardware tests.
Xinyu2ML: Xinyu Yang, known on X as @Xinyu2ML, is an engineer and architect at Moonshot AI (Kimi) with a background as a PhD from Carnegie Mellon University. They focus on scaling open frontier intelligence models. The quoted post from this user directly highlights Kimi K3’s ability to write custom GPU kernels as demonstrated in the KernelBench evaluation.

Kernel Generation: KernelBench evaluates AI models on producing custom GPU code that must both compile correctly and deliver measurable performance gains over reference implementations on actual hardware.
AI Coding Benchmarks: Recent evaluations position Kimi K3 as highly competitive with leading models like Claude Opus 4.8 specifically in advanced coding tasks involving low-level optimizations.

AI tools built by Emerald Force

Built and supported by Emerald Force.

You might also like