16 November 2026

Several labs report faster AI model serving on AMD chips and new software

First reported

Latent Space ran this on .

  • Reports say a vLLM setup with FlyDSL kernels ran MiniMax M3.1 faster on AMD's MI355X than on Nvidia's Blackwell chips.
  • Baseten, a company that hosts AI models for customers, built a new engine in Rust and CUDA for DeepSeek V4.1-Flash.

How it was covered