Hardware

Ryzen AI Max Strix Halo APU runs 51 LLMs ported by AI agents in four days

By One Percent LowsCovered by 2 outlets
★ Add us as a preferred source on Google

TL;DR

AMD engineers used AI coding agents and HRX Loom version 0.1 between November 6 and 9 to bring 51 trending Hugging Face models to Ryzen AI Max Strix Halo hardware, achieving a 1.5x speed-up over llama.cpp on long prompts.

AMD engineers Jeremy Fowers and Ben Vanik used AI coding agents and the HRX Loom compiler to port 51 trending language models to client hardware in four days. The project ran between November 6 and 9 across six DevLab machines and one local system powered by Ryzen AI Max Strix Halo hardware.

The target list covered the top 50 trending builds on Hugging Face as of November 6 alongside Qwen3.8-Flash-Next, spanning 19 architectures. The models included Qwen variants, Gemma 4, Ornith builds up to 122 billion parameters, MiniCPM-V, and an NVIDIA Nemotron build. Using tools like Claude Code, Codex, and Pi, a median of two subagents worked concurrently while the main agent peaked at 16.

Benchmarks comparing the generated kernel files against llama.cpp showed a median 1.5x speed-up for 32K-token prompt processing against the ROCm back-end, with the spread reaching up to 2.1x. Performance reached roughly 1.9x faster than the Vulkan path in llama.cpp.

HRX Loom version 0.1 remains experimental with a tech-preview label. AMD notes the four-day turnaround serves as a showcase rather than a support promise, with work centered on systems such as the Ryzen AI Max-based Framework Desktop.

More on AMDAI in gaming and hardware