AMD Had AI Agents Port 51 Trending LLMs to Strix Halo in Four Days, and They Beat llama.cpp on Long Prompts

AMD engineers Jeremy Fowers and Ben Vanik used AI coding agents to port and optimize 51 trending large language models across 19 architectures to Ryzen AI Max "Strix Halo" client GPUs in roughly four days, between November 6 and 9, using AMD's HRX Loom compiler and loom-skills tooling. In benchmarks against llama.cpp on the same hardware, the ports delivered a median 1.5x speed-up on 32K-token prompt processing versus llama.cpp's ROCm back-end, with Phoronix reporting a spread of roughly 1.5x to 2.1x and about 1.9x versus llama.cpp's Vulkan path. Loom, still labeled tech-preview with v0.1 released recently, compiles model-specific kernels written in a C++ subset or Loom IR that AMD calls "agent-friendly," letting agents author and tune code outside a monolithic inference engine.

An AMD team pointed a squad of AI coding agents at the 50 hottest LLMs on Hugging Face and had them running on Ryzen AI Max “Strix Halo” in about four days, with long-prompt processing typically 1.5x faster than llama.cpp’s ROCm back-end. Getting a brand-new model to run well on anything other than an NVIDIA GPU has usually meant waiting. Waiting for llama.cpp to add the architecture, for someone to write decent kernels, for the quantized builds to show up. AMD is now trying to shrink that wait with its own compiler work plus a lot of agent labor, and the first public experiment is more interesting than the usual demo. As Phoronix reports https://www.phoronix.com/news/AMD-HRX-Loom-Experiment , AMD engineers Jeremy Fowers and Ben Vanik ran a campaign that took the top 50 trending LLMs on Hugging Face as of November 6, added Qwen3.8-Flash-Next on top, and ported and optimized all of them for AMD client GPUs using HRX Loom. AMD’s results dashboard https://loom-fun.github.io/loom-skills/ counts 51 models across 19 architectures, finished over roughly four days of work between November 6 and 9. What Loom actually is Loom sits inside AMD’s HRX stack, and the project documentation https://rocm.github.io/hrx-system/loom/ describes it as a language and compiler for native programs on CPUs, GPUs and NPUs. Programs can be written in a subset of C++ or in Loom IR, which AMD openly calls “agent-friendly”, and the compiler spits out ELF, HSACO, SPIR-V, WebAssembly and other targets. Models ship as small packages .loom source or bytecode, weights and a JSON descriptor served by loom-serve, which can load a new model or swap an implementation without a restart. It runs on Windows and Linux and still carries a tech-preview label; v0.1 only went out recently. That design is the whole trick. When the model-specific code lives in compact, readable kernel files instead of deep inside an inference engine, an AI agent has something tractable to write, test and tune. How the experiment ran The team packaged that know-how as loom-skills https://github.com/loom-fun/loom-skills , a set of skills that equip Claude Code, Codex or the Pi agent to author and serve optimized model programs. The hardware was modest: six Strix Halo machines in AMD’s DevLab plus one local Strix Halo box. According to the methodology notes, the orchestrating agent ran a median of two subagents at a time, peaking at 16, with up to six DevLab jobs in flight. The model list is a fair snapshot of what people actually download right now: Qwen3.5, 3.6 and 3.8 variants, Gemma 4, several Ornith builds up to 122B parameters, MiniCPM-V and an NVIDIA Nemotron model among them. Each port gets a Hugging Face page with serving instructions and the generated .loom kernel files. The numbers Every port was benchmarked against llama.cpp on the same hardware, using greedy, thinking-off chat completions with a unique nonce at the start of each prompt so prefix caching could not flatter either engine. The headline figure is a median 1.5x speed-up for 32K-token prompt processing versus llama.cpp’s ROCm back-end. Phoronix puts the spread at roughly 1.5x to 2.1x, and about 1.9x against llama.cpp’s Vulkan path. The dashboard also tracks decode speed and four-stream throughput per model, so you can see where Loom wins and where it does not. Prefill is exactly where Strix Halo-class machines hurt. With up to 128GB of unified memory they can hold genuinely large models, but chewing through a long document or a codebase before the first token appears is often the painful wait. A 1.5x gain there is something you feel. Why it matters The caveats are real. Loom is experimental, the work is focused on and tested with Strix Halo systems such as the Ryzen AI Max-based Framework Desktop, and four days of agent time for 51 models is a showcase, not a support promise. Still, AMD’s problem in local AI has rarely been the silicon. It has been how long new models take to run properly on it. If agents plus an agent-friendly compiler can turn that lag into days instead of months, that changes the pitch for AMD’s big APUs more than another SKU would.