Moonshot AI открыла исходный код FlashKDA: ускорение prefill до 2,22 раза на H20

Compiled on May 1, 2026 ## Key Highlights The week’s most significant technical advancement is the open-sourcing of FlashKDA by Moonshot AI. This implementation offers high-performance kernels for Kimi Delta Attention that plug directly into the flash-linear-attention ecosystem, specifically leveraging the CUTLASS framework. By introducing variable-length batching and benchmarking against H20 hardware, Moonshot is addressing a critical need for inference efficiency. This move suggests that while model scaling has been a priority, optimizing the underlying attention mechanisms for specific hardware is now a key driver for competitive advantage in the open-source community. [...] ## Papers to Read ### 1. Moonshot AI Open-Sources FlashKDA: CUTLASS Kernels for Kimi Delta Attention with Variable-Length Batching and H20 Benchmarks Source: MarkTechPost Moonshot AI releases FlashKDA, a high-performance implementation of Kimi Delta Attention that plugs directly into the flash-linear-attent

Agree & Join LinkedIn By clicking Continue to join or sign in, you agree to LinkedIn’s User Agreement, Privacy Policy, and Cookie Policy. # FlashKDA: High-Performance Attention Kernels Open-Sourced View organization page for Kimi (Moonshot AI) 46,353 followers We've open-sourced FlashKDA, our high-performance CUTLASS-based implementation of Kimi Delta Attention kernels. It delivers 1.72×–2.22× prefill speedup over the flash-linear-attention baseline on H20, and works as a drop-in backend for flash-linear-attention. Explore on GitHub: GitHub - MoonshotAI/FlashKDA: FlashKDA: high-performance Kimi Delta Attention kernels Chris Qin, graphic [...] Chris Qin, graphic Impressive numbers, especially with a drop-in backend. Kernel improvements like this directly expand the practical context and throughput envelope. Curious how performance scales across sequence lengths and batch sizes on H20. Simon Jakobs, graphic Wow thanks! Will use it when serving Kimi K3 fully EU hosted on meliou

By Asif Razzaq - The team behind Kimi.ai (Moonshot AI) just made a significant contribution to the open-source AI infrastructure space. The research team has made a significant contribution to the open-source AI infrastructure space. They released FlashKDA (Flash Kimi Delta Attention), a high-performance CUTLASS-based kernel implementation of the Kimi Delta Attention (KDA) mechanism. The FlashKDA library is available on GitHub under an MIT license. It delivers prefill speedups of 1.72× to 2.22× over the `flash-linear-attention` baseline on NVIDIA H20 GPUs, and works as a drop-in backend for the popular `flash-linear-attention` library. ### What Is Kimi Delta Attention, and Why Does It Matter? [...] FlashKDA is Moonshot AI’s open-source CUTLASS-based CUDA kernel for Kimi Delta Attention (KDA), delivering 1.72×–2.22× prefill speedup over the `flash-linear-attention` baseline on NVIDIA H20 GPUs. KDA extends Gated DeltaNet with fine-grained, channel-wise gating — it’s the core attenti

By Asif Razzaq - The team behind Kimi.ai (Moonshot AI) just made a significant contribution to the open-source AI infrastructure space. The research team has made a significant contribution to the open-source AI infrastructure space. They released FlashKDA (Flash Kimi Delta Attention), a high-performance CUTLASS-based kernel implementation of the Kimi Delta Attention (KDA) mechanism. The FlashKDA library is available on GitHub under an MIT license. It delivers prefill speedups of 1.72× to 2.22× over the `flash-linear-attention` baseline on NVIDIA H20 GPUs, and works as a drop-in backend for the popular `flash-linear-attention` library. ### What Is Kimi Delta Attention, and Why Does It Matter? [...] FlashKDA is Moonshot AI’s open-source CUTLASS-based CUDA kernel for Kimi Delta Attention (KDA), delivering 1.72×–2.22× prefill speedup over the `flash-linear-attention` baseline on NVIDIA H20 GPUs. KDA extends Gated DeltaNet with fine-grained, channel-wise gating — it’s the core attenti

# Thread 595 views sung.kim.mw's profile picture sung.kim.mw Moonshot AI's FlashKDA A high-performance CUTLASS-based implementation of Kimi Delta Attention kernels. Achieves 1.72×–2.22× prefill speedup over the flash-linear-attention baseline on H20, and works as a drop-in backend for flash-linear-attention. github.com/Moons… GitHub - MoonshotAI/FlashKDA: FlashKDA: high-performance Kimi Delta Attention kernels github.com GitHub - MoonshotAI/FlashKDA: FlashKDA: high-performance Kimi Delta Attention kernels 15 1 1 1 Log in or sign up for ThreadsSee what people are talking about and join the conversation. Log in with username instead

Log inSign up ## Post user avatar Kimi.ai @Kimi\_Moonshot We've open-sourced FlashKDA, our high-performance CUTLASS-based implementation of Kimi Delta Attention kernels. It delivers 1.72×–2.22× prefill speedup over the flash-linear-attention baseline on H20, and works as a drop-in backend for flash-linear-attention. Explore on GitHub: github.com/MoonshotAI/Fla… GitHub - MoonshotAI/FlashKDA: FlashKDA: high-performance Kimi Delta Attention kernelsFrom github.com 3:25 PM · Jul 27, 2026168.8KViews user avatar Louis（没 Fable 5 用我就要发狂了） @0x\_Yao Jul 27 Tell me, who's the real hero? GIF 5.4K user avatar Kirolos Basalos @KeroBasalos Jul 27 [...] Kirolos Basalos @KeroBasalos Jul 27 I FUCKING love you guys. Ong, fuck “open” AI, definitely fuck anthropic, kinda fuck meta. My only hope is in Gemini, Kimi and Qwen. Y’all are beautiful goddamnit. 3.1K user avatar Bing Xu @bingxu\_ Jul 27 We’ve open-sourced PTX KDA, our fu