AMD's HRX Loom: Porting 50 AI Models to Client GPUs in Days
AMD is tackling one of the biggest hurdles in local AI adoption: the software gap. Through an experimental project, the company has successfully ported 50 popular Hugging Face models from 18 different architectures to client GPUs in just a few days. This was achieved using HRX Loom combined with an agentic development workflow, where AI tools handle much of the heavy lifting of the porting process. While the results show significant performance gains in specific scenarios, the project remains in an early, experimental stage.
The real challenge for local AI hardware isn't just raw compute power or memory capacity; it's the software ecosystem. Models require specific kernels, runtimes, and optimized implementations to run efficiently on particular hardware. For less common or brand-new models, this optimization process can take a significant amount of time. AMD is attempting to automate this overhead with Loom, aiming to bridge the gap between hardware availability and software readiness.
Automating the Porting Process with AI Agents
The current Loom version 0.1 includes compilers, runtimes, command-line tools, and Python interfaces for x86-64 systems running both Linux and Windows. While the project is designed to support CPU and NPU execution, the current focus is heavily centered on AMD GPUs and client-side local model execution. To achieve these rapid results, AMD engineers combined Loom with so-called 'loom-skills.' These provide AI agents—such as Codex, Claude Code, or Pi—with the technical information and workflows necessary to generate, adapt, and test model programs. Instead of manually implementing every single kernel, these agents take over repetitive porting tasks and then test the results against real hardware.
The experiment successfully covered 50 Hugging Face models spanning 18 different architectures. This variety is a key detail; rather than simply porting 50 variations of a single base model, the project demonstrated the ability to handle diverse model structures.
Performance Gains and Benchmarks
A concrete example of this workflow is the publicly available Loom implementation of Qwen3.8-Flash-Next. This package includes Loom programs and a model definition for an unmodified quantized GGUF checkpoint, which was generated on a Ryzen AI Max+ 395. The model stack can be launched using AMDGPU as the target device and provides an OpenAI-compatible chat interface. AMD and the participating developers report performance improvements over llama.cpp. Specifically, when processing a 32K prompt, they reported throughput values approximately 1.5 to 2.1 times higher than the llama.cpp ROCm backend, and about 1.9 times higher than the Vulkan backend. It is important to note that these figures come from an experimental project and are not part of a large-scale, independent benchmark.
Because of the experimental nature of the tests, these numbers should not be used to claim a general performance advantage for Loom over llama.cpp. Factors such as the specific model, quantization method, context length, batch size, hardware, and the kernels used all heavily influence throughput. The measurements simply demonstrate that the new stack is capable of being highly competitive in these specific scenarios.
Closing the Ecosystem Gap
The most significant implication of this project is the speed of porting. If AMD can indeed bring dozens of models to new client hardware within a few days, it could mitigate a long-standing disadvantage: the software lag. Currently, new GPU architectures often arrive before optimized software for smaller or niche models is available. This is a critical area for AMD, as NVIDIA's CUDA ecosystem has benefited from years of massive software development. While an agent-based compiler approach won't replace the CUDA ecosystem overnight, it could drastically reduce the time between a model's release and a functional AMD implementation.
However, HRX Loom remains strictly experimental. Even the publicly documented model packages may require development builds, and their interfaces or installation methods are subject to change. For instance, the Qwen3.8 example currently requires a more recent development version than the official Loom 0.1 release available on PyPI.