AMD's Loom vs llama.cpp on a Strix Halo laptop, on Windows

Running Qwen3.8 27B and two Ornith 1.5 models on AMD's early Loom engine and on llama.cpp with Vulkan, on an ASUS ROG Flow Z13: faster on the dense models, slower on the mixture-of-experts one.

5 min read

Loom (hrx-loom 0.1.0) is an early inference engine from AMD’s ROCm organization. It only runs models that have been ported to it, each port pins one exact GGUF file, and it ships its own OpenAI-compatible server. I wanted to know whether it beats llama.cpp on my Strix Halo laptop, on real work rather than a synthetic benchmark. This is one of four posts from my local LLMs page, which also explains the stack and the model types if any of the terms here are new.

The short version: Loom was faster on the two dense models and slower on the mixture-of-experts one.

The setup

Everything ran on an ASUS ROG Flow Z13 with a Ryzen AI Max+ 395 (Strix Halo) and its Radeon 8060S integrated GPU, with 96 GB of its 128 GB set aside for the GPU, on Windows 11. The llama.cpp side was b12387 with the Vulkan backend.

The work was the three unattended code repair projects. Each run gives the model a small broken JavaScript project, a short task and a few file tools (list, read, write, edit, run tests), then lets it loop with nobody to answer questions. Afterwards the harness re-runs the original tests and a set of hidden checks the model never saw. A pass means every visible test and every hidden check passed.

Qwen3.8 27B

First Qwen3.8 27B, with no thinking cap:

Projectllama.cpp Vulkanllama.cpp + MTPLoom
Sales summary273 s102 s*125 s
Interval mergenot run49 s49 s
Dispatch queue493 s193 s*154 s
Output tokens per second, whole run11.6–11.923.1–26.425.2–25.7

* thinking capped at 4,000 tokens. Every run passed.

Loom gave about twice the throughput of plain llama.cpp on Vulkan, and was roughly level with llama.cpp + MTP. One caveat: Loom only runs the exact GGUF file its model card pins, so the llama.cpp runs used a different build of the same Q4_K_M quantisation.

Two Ornith 1.5 models

I then ran two Ornith 1.5 models the same way, this time with the identical GGUF file on both engines and no thinking cap. Speed is output tokens per second over the whole repair run:

ModelEngineSpeedSales summaryInterval mergeDispatch queue
Ornith 1.5 9B Q4_K_MLoom50–51pass, 99 spass, 75 sfail (1/8, 8/32), 491 s
Ornith 1.5 9B Q4_K_Mllama.cpp Vulkan33–34pass, 710 spass, 107 sfail (1/8, 8/32), 249 s
Ornith 1.5 35B-A3B Q4_K_MLoom40–45pass, 124 spass, 68 spass, 240 s
Ornith 1.5 35B-A3B Q4_K_Mllama.cpp Vulkan54–56pass, 78 spass, 41 sfail (8/8, 31/32), 123 s
  • On the dense 9B, Loom was about 1.5× faster. The 9B itself is too small for the dispatch queue: on both engines it left the code no better than it started.
  • On the mixture-of-experts 35B-A3B, Loom was slower, at about 0.75× llama.cpp’s throughput, so every repair took longer. That is the opposite of what I saw on the two dense models.
  • Loom read prompts faster on both. A single 4,000-token prompt went through at roughly 1,500 tok/s on the 9B and 2,700 on the 35B-A3B, against llama-bench’s 858 and 1,210 for llama.cpp. Those are different measurements, so treat the ratio as rough.

Two things to keep in mind. Loom’s Ornith packages include a small drafter model for speculative decoding and llama.cpp ran without one, so the comparison favours Loom if anything. And the times swing with how much the model chooses to write: the 9B’s 710 seconds on llama.cpp was 14 passes and 23,500 tokens, against 5 passes and 4,900 tokens on Loom.

The model I couldn’t test

I couldn’t test the model I most wanted to. Loom only runs models that have been ported to it, and Flash-Next is ported only as a 111 GB quantisation, which doesn’t fit in the Z13’s 96 GB. The released 0.1.0 package also can’t load the current Flash-Next programs. I’ve asked about smaller quants.

Also worth knowing: the speedups on the public Loom leaderboard are measured on Linux against llama.cpp’s ROCm backend, not on Windows against Vulkan, so my numbers aren’t comparable with theirs.

Caveats

  • Single runs. Every result above is one attempt. A different seed could pass or fail differently, especially on the borderline ones.
  • Windows only, so far. Linux, with ROCm, is next, and that’s where Loom’s own comparisons are made.
  • Model settings differ between columns where noted (MTP on or off, thinking capped or not).

Written by

Farez Vadsaria

Senior engineer on AI voice services at Medsender, previously identity & auth at Warner Bros and email infrastructure at Amazon. Builds free AI side projects with coding agents, and teaches how they work.

Get the next post when it ships.