FrontierThe story, in brief
Qwen Team Releases FlashQLA: a High-Performance Linear Attention Kernel Library That Achieves Up to 3× Speedup on NVIDIA Hopper GPUs
3× speedup. That's what Qwen's new FlashQLA kernel achieves on Hopper GPUs—reshaping inference economics for agentic AI.
The KeyNews take
Why it matters
Qwen's FlashQLA kernel library demonstrates a critical efficiency play in model inference, directly impacting deployment costs and latency for both pretraining and edge inference. Linear attention optimizations like this are becoming table-stakes for competitive model serving.
The key facts
6 to knowFlashQLA achieves up to 3× speedup on NVIDIA Hopper GPUs
Targets Gated Delta Network (GDN) Chunked Prefill
Optimizes both forward and backward passes
Designed for large-scale pretraining and edge-side agentic inference
Released by QwenLM team
Published Apr 29, 2026
Go to the source
MarkTechPostmarktechpost.com
Publisher excerpt: The QwenLM team has released FlashQLA, a new kernel library that dramatically accelerates the forward and backward passes of Gated Delta Network (GDN) Chunked Prefill, targeting both large-scale pretraining and edge-side agentic inference scenarios. The post Qwen Team Releases FlashQLA: a…