An AMD team has ported 51 popular language models to Ryzen AI Max GPUs, known by the codename Strix Halo, in just four days — and not with a huge team of engineers, but with AI agents writing and optimizing the code. The experiment, led by engineers Jeremy Fowers and Ben Vanik, used the HRX Loom compiler and a set of skills for coding agents, loom-skills.
The starting point was the list of the 50 most popular LLMs on Hugging Face as of November 6, 2026, plus Qwen3.8-Flash-Next. In total, 51 models across 19 architectures, including Qwen3.5, Qwen3.6 and Qwen3.8 variants, Gemma 4, Ornith models of up to 122 billion parameters, MiniCPM-V and even an NVIDIA Nemotron model. Each port got a Hugging Face page with usage instructions and the generated .loom kernel files.
The numbers
Every model was benchmarked against llama.cpp on the same hardware, using greedy chat completions with a unique nonce at the start of each prompt, so prefix caching could not favor either side. The headline result: a median 1.5x speedup on 32K-token prompt processing (prefill) versus llama.cpp's ROCm backend. The spread ran from about 1.5x to 2.1x, and about 1.9x against llama.cpp's Vulkan backend. The results dashboard also tracks decode speed and four-stream throughput per model.
The hardware was modest: six Strix Halo machines in AMD's DevLab plus one local box, all Ryzen AI MAX+ 395 with Radeon 8060S graphics, running Ubuntu 24.04. The orchestrating agent ran a median of two subagents at a time, peaking at 16.
What HRX Loom is
Loom is a language and compiler for native programs on CPUs, GPUs and NPUs, part of AMD's HRX (HIP Runtime Extended) stack. Programs can be written in a subset of C++ or in Loom IR, which AMD itself describes as "agent-friendly", and the compiler emits ELF, HSACO, SPIR-V and WebAssembly, among other targets. Models ship as packages with .loom source or bytecode, weights and a JSON descriptor, served by loom-serve — which can load a new model or swap an implementation without restarting the server. Version 0.1 of the SDK was released recently and the project is still in tech preview, for Windows and Linux.
The trick is in the architecture: when the model-specific code lives in compact, readable kernel files instead of being buried deep inside an inference engine, an AI agent has something tractable to write, test and tune. The loom-skills package, published on GitHub, equips agents such as Claude Code, Codex and Pi to author and serve optimized model programs.
Why it matters
AMD's problem in local AI has rarely been the silicon: Strix Halo chips carry up to 128 GB of unified memory, enough to run large models on a single machine. The bottleneck has always been how long it takes new models to run well on that hardware. If agents plus a compiler built for them can shrink that wait from months to days, the case for AMD's APUs changes shape — and the catalog of already-ported models is public on Hugging Face for anyone to try. The same kind of hardware is already reaching the market: the BOSGAME M5 Max packs 192 GB of unified memory with the Ryzen AI Max+ PRO 495.
There are caveats, and they matter: Loom is experimental, the work was focused on and tested with Strix Halo systems such as the Framework Desktop, and four days of agent work is a showcase, not a support promise. The gains also vary by model — the dashboard shows where Loom wins and where it does not.
Sources
- Phoronix: AMD Leverages AI + HRX Loom In Experiment For Quickly Porting 50 LLMs To Client GPUs
- Experiment results dashboard (loom-fun)
- Official Loom documentation (AMD ROCm)
- loom-skills on GitHub
- Ported models on Hugging Face (loom-fun)
- Hardware Busters: AMD Had AI Agents Port 51 Trending LLMs to Strix Halo in Four Days
Image credit: AMD, via Phoronix.
Transparency: this article was produced with the help of artificial intelligence, cross-checked against public posts on X and sources on the internet. Please consult the original sources to confirm the information.
By Muse
