Elliot Arledge

@elliotarledge

@infinity_ai_ http://kernelbench.com | made the 12 hr CUDA course | http://elliotarledge.com/books | @shipfr8 alum

following
followers

DeepSeek V4.1 Flash reportedly ranks first of 23 on MegaQwen decode

A post reports all four handwritten CUDA kernels passed on an RTX Pro 6000; Sonnet 5.5 reached 8.00% of roofline.

2 Sources, 1d ago, first seen 1d ago

DeepSeek V4.1 Flash (1017) on KernelBench-CUDA, four problems vs the September checkpoint and the board leader

Husky is said to run one model on Mac and iPhone without Wi-Fi

A post describes Husky as a model-specific inference engine. A follow-up says its Metal kernels are tailored to the Woof model, with the goal of beating Apple's MLX.

5 Sources, 28d ago, updated 28d ago

Grok 4.7 launches in Cursor, Grok Build and the Grok API

SpaceXAI calls it an improvement over Grok 4.6 at the same price and speed. Artificial Analysis reports better coding results, but more than double the output-token use in a separate evaluation.

70 Sources, 26d ago, updated 26d ago

DeepSeek V4.1 Flash’s reported fourth-place sparse-attention benchmark ranking

A KernelBench-CUDA tester reports runtimes of 0.059–0.736 milliseconds across six test shapes on an RTX PRO 6000 and attributes the entire gap to the top three to unnecessary computation.

4 Sources, 39d ago, first seen 39d ago

Featured in

DeepSeek V4.1 Flash reportedly ranks first of 23 on MegaQwen decode

Husky is said to run one model on Mac and iPhone without Wi-Fi

Grok 4.7 launches in Cursor, Grok Build and the Grok API

DeepSeek V4.1 Flash’s reported fourth-place sparse-attention benchmark ranking

Featured in

DeepSeek V4.1 Flash reportedly ranks first of 23 on MegaQwen decode

Husky is said to run one model on Mac and iPhone without Wi-Fi

Grok 4.7 launches in Cursor, Grok Build and the Grok API

DeepSeek V4.1 Flash’s reported fourth-place sparse-attention benchmark ranking