https://preview.redd.it/eq2nojgspsug1.png?width=780&format=png&auto=webp&s=4e0517c673e06dd1995f32b89363c75315dfffb9
So I wanted a portable 13 inch laptop that can be a little LLM monster when needed, Asus did an amazing job with their new 2026 PX13 laptopn powered by strixhalo 128G unified memeory APU
I made benchmark automation system for the amazing toolboxs repo here:
https://github.com/kyuz0/amd-strix-halo-toolboxes
This repo gives you multiple ready to use llamacpp builds with rocm and vulkan
my script is setting the power profile to either (power saving or high performance) then benchmark with llama-bench all the provided gguf with 3 diffrent llama backend (vulkan/rocm nightly/amdvlk)
the overall benchmark for 25 models (varies from 4B to 120B) with all diffrent backends and powerprofils, this took almost 12 hours with average time 4 ~ 5 minutes per run for each model at each configuration
side note: I tested multiple "heretic/hauhau versions" of the mainstream model because I found they are much efficient at thinking process and I saw littel increase in their coding performance comparing to original ones (with some drop in transaltions tasks)
Here is the visualized leaderboard
Token Generation leaderboard
Prompt Processing leaderboard
for power profile power saving I saw consumption near 40 watt and for performance it varies from 60 - 77 watt
------------
llama-bench ProArt PX13 HN7306EAC with strix halo toolboxes
- Machine model:
ProArt PX13 HN7306EAC - CPU:
AMD RYZEN AI MAX+ 395 w/ Radeon 8060S - Architecture:
x86_64 - Kernel:
7.0.0-rc7-2-cachyos-rc - OS:
CachyOS n/a - OS Version:
n/a - Toolboxes:
['llama-rocm7-nightlies', 'llama-vulkan-amdvlk', 'llama-vulkan-radv'] - Mode:
medium - Power Profiles:
['performance', 'power-saver'] - Prompt tokens:
1024,4096,8192,16384 - Generation tokens:
512,2048 - Repetitions:
1
Leaderboard (sorted by Token Generation/Second)
| Rank | Model | Best Gen Backend | Power Profile | Prompt/Gen Tokens (Gen) | Best Gen TPS | Best Prompt Backend | Prompt/Gen Tokens (Prompt) | Best Prompt TPS |
| 1 | Marco-Nano-Instruct.Q8_0.gguf | llama-vulkan-radv | Performance | 512 | 211.325 | llama-vulkan-radv | 1024 | 4296.133 |
| 2 | Marco-Mini-Instruct.Q8_0.gguf | llama-vulkan-radv | Performance | 512 | 165.874 | llama-vulkan-radv | 1024 | 2329.999 |
| 3 | OpenAI-20B-NEO-CODEPlus-Uncensored-IQ4_NL.gguf | llama-vulkan-radv | Performance | 512 | 86.033 | llama-rocm7-nightlies | 1024 | 1347.876 |
| 4 | gpt-oss-20b-Derestricted-MXFP4_MOE.gguf | llama-vulkan-radv | Performance | 512 | 74.471 | llama-rocm7-nightlies | 1024 | 1317.919 |
| 5 | gpt-oss-20b-heretic.MXFP4_MOE.gguf | llama-vulkan-radv | Performance | 512 | 74.356 | llama-vulkan-radv | 1024 | 1323.742 |
| 6 | Qwen3.5-35B-A3B-Uncensored-HauhauCS-Aggressive-Q4_K_M.gguf | llama-vulkan-amdvlk | Performance | 512 | 69.059 | llama-vulkan-radv | 1024 | 917.500 |
| 7 | Qwen3.5-35B-A3B-heretic.Q4_K_M.gguf | llama-vulkan-amdvlk | Performance | 512 | 69.001 | llama-vulkan-radv | 1024 | 928.552 |
| 8 | LFM2-24B-A2B-Q8_0.gguf | llama-vulkan-amdvlk | Power Saver | 512 | 60.739 | llama-rocm7-nightlies | 1024 | 1456.713 |
| 9 | Qwen3.5-35B-A3B-Q4_K_M.gguf | llama-vulkan-amdvlk | Power Saver | 512 | 59.614 | llama-rocm7-nightlies | 1024 | 911.428 |
| 10 | Qwen3.5-4B-Uncensored-HauhauCS-Aggressive-Q4_K_M.gguf | llama-vulkan-radv | Performance | 512 | 59.263 | llama-vulkan-radv | 1024 | 1716.063 |
| 11 | Qwen3.5-4B-UD-Q4_K_XL-unsloth-v2.gguf | llama-vulkan-radv | Performance | 512 | 56.642 | llama-vulkan-radv | 4096 | 1600.179 |
| 12 | gemma-4-26B-A4B-it-UD-Q3_K_M.gguf | llama-vulkan-radv | Performance | 512 | 55.191 | llama-rocm7-nightlies | 1024 | 1044.901 |
| 13 | gemma-4-26B-A4B-it-UD-IQ4_XS.gguf | llama-vulkan-radv | Performance | 512 | 52.416 | llama-rocm7-nightlies | 1024 | 1510.919 |
| 14 | bartwoski_Qwen3.5-35B-A3B-Q4_K_M.gguf | llama-vulkan-amdvlk | Power Saver | 512 | 51.307 | llama-rocm7-nightlies | 1024 | 783.849 |
| 15 | gemma-4-26B-A4B-it-UD-Q4_K_XL (1).gguf | llama-vulkan-radv | Performance | 512 | 49.469 | llama-rocm7-nightlies | 1024 | 1620.560 |
| 16 | Qwen3-Coder-Next-UD-IQ1_M.gguf | llama-vulkan-radv | Power Saver | 512 | 48.834 | llama-vulkan-radv | 1024 | 472.070 |
| 17 | Qwen3.5-35B-A3B-UD-Q4_K_XL-unsloth-v2.gguf | llama-vulkan-amdvlk | Power Saver | 512 | 46.992 | llama-rocm7-nightlies | 1024 | 1009.841 |
| 18 | bartwoski_Qwen3-Coder-Next-IQ4_XS.gguf | llama-vulkan-radv | Power Saver | 512 | 41.375 | llama-vulkan-radv | 1024 | 615.839 |
| 19 | kldzj_gpt-oss-120b-heretic-v2-MXFP4_MOE-00001-of-00002.gguf | llama-rocm7-nightlies | Power Saver | 512 | 40.004 | llama-vulkan-radv | 1024 | 432.180 |
| 20 | Qwen_Qwen3-Coder-Next-IQ4_XS.gguf | llama-vulkan-radv | Power Saver | 0/2048 | 39.801 | llama-vulkan-radv | 1024 | 621.813 |
| 21 | Qwen3.5-9B-Uncensored-HauhauCS-Aggressive-Q4_K_M.gguf | llama-vulkan-radv | Performance | 512 | 36.393 | llama-rocm7-nightlies | 1024 | 953.875 |
| 22 | Qwen3.5-122B-A10B-Uncensored-HauhauCS-Aggressive-IQ3_XXS.gguf | llama-vulkan-radv | Power Saver | 512 | 27.562 | llama-rocm7-nightlies | 1024 | 186.736 |
| 23 | omnicoder-2-9b-q8_0.gguf | llama-vulkan-radv | Performance | 512 | 23.944 | llama-rocm7-nightlies | 1024 | 986.071 |
| 24 | bartwoski_Qwen3.5-122B-A10B-IQ3_XXS-00001-of-00002.gguf | llama-vulkan-radv | Power Saver | 512 | 23.206 | llama-rocm7-nightlies | 1024 | 234.785 |
| 25 | unsloth-Qwen3.5-122B-A10B-UD-IQ3_XXS.gguf | llama-vulkan-radv | Power Saver | 512 | 20.771 | llama-rocm7-nightlies | 1024 | 194.398 |
Leaderboard (sorted by Prompt Processing T/Second)
| Rank | Model | Best Gen Backend | Power Profile | Prompt/Gen Tokens (Gen) | Best Gen TPS | Best Prompt Backend | Prompt/Gen Tokens (Prompt) | Best Prompt TPS |
| 1 | Marco-Nano-Instruct.Q8_0.gguf | llama-vulkan-radv | Performance | 512 | 211.325 | llama-vulkan-radv | 1024 | 4296.133 |
| 2 | Marco-Mini-Instruct.Q8_0.gguf | llama-vulkan-radv | Performance | 512 | 165.874 | llama-vulkan-radv | 1024 | 2329.999 |
| 3 | Qwen3.5-4B-Uncensored-HauhauCS-Aggressive-Q4_K_M.gguf | llama-vulkan-radv | Performance | 512 | 59.263 | llama-vulkan-radv | 1024 | 1716.063 |
| 4 | gemma-4-26B-A4B-it-UD-Q4_K_XL (1).gguf | llama-vulkan-radv | Performance | 512 | 49.469 | llama-rocm7-nightlies | 1024 | 1620.560 |
| 5 | Qwen3.5-4B-UD-Q4_K_XL-unsloth-v2.gguf | llama-vulkan-radv | Performance | 512 | 56.642 | llama-vulkan-radv | 4096 | 1600.179 |
| 6 | gemma-4-26B-A4B-it-UD-IQ4_XS.gguf | llama-vulkan-radv | Performance | 512 | 52.416 | llama-rocm7-nightlies | 1024 | 1510.919 |
| 7 | LFM2-24B-A2B-Q8_0.gguf | llama-vulkan-amdvlk | Power Saver | 512 | 60.739 | llama-rocm7-nightlies | 1024 | 1456.713 |
| 8 | OpenAI-20B-NEO-CODEPlus-Uncensored-IQ4_NL.gguf | llama-vulkan-radv | Performance | 512 | 86.033 | llama-rocm7-nightlies | 1024 | 1347.876 |
| 9 | gpt-oss-20b-heretic.MXFP4_MOE.gguf | llama-vulkan-radv | Performance | 512 | 74.356 | llama-vulkan-radv | 1024 | 1323.742 |
| 10 | gpt-oss-20b-Derestricted-MXFP4_MOE.gguf | llama-vulkan-radv | Performance | 512 | 74.471 | llama-rocm7-nightlies | 1024 | 1317.919 |
| 11 | gemma-4-26B-A4B-it-UD-Q3_K_M.gguf | llama-vulkan-radv | Performance | 512 | 55.191 | llama-rocm7-nightlies | 1024 | 1044.901 |
| 12 | Qwen3.5-35B-A3B-UD-Q4_K_XL-unsloth-v2.gguf | llama-vulkan-amdvlk | Power Saver | 512 | 46.992 | llama-rocm7-nightlies | 1024 | 1009.841 |
| 13 | omnicoder-2-9b-q8_0.gguf | llama-vulkan-radv | Performance | 512 | 23.944 | llama-rocm7-nightlies | 1024 | 986.071 |
| 14 | Qwen3.5-9B-Uncensored-HauhauCS-Aggressive-Q4_K_M.gguf | llama-vulkan-radv | Performance | 512 | 36.393 | llama-rocm7-nightlies | 1024 | 953.875 |
| 15 | Qwen3.5-35B-A3B-heretic.Q4_K_M.gguf | llama-vulkan-amdvlk | Performance | 512 | 69.001 | llama-vulkan-radv | 1024 | 928.552 |
| 16 | Qwen3.5-35B-A3B-Uncensored-HauhauCS-Aggressive-Q4_K_M.gguf | llama-vulkan-amdvlk | Performance | 512 | 69.059 | llama-vulkan-radv | 1024 | 917.500 |
| 17 | Qwen3.5-35B-A3B-Q4_K_M.gguf | llama-vulkan-amdvlk | Power Saver | 512 | 59.614 | llama-rocm7-nightlies | 1024 | 911.428 |
| 18 | bartwoski_Qwen3.5-35B-A3B-Q4_K_M.gguf | llama-vulkan-amdvlk | Power Saver | 512 | 51.307 | llama-rocm7-nightlies | 1024 | 783.849 |
| 19 | Qwen_Qwen3-Coder-Next-IQ4_XS.gguf | llama-vulkan-radv | Power Saver | 0/2048 | 39.801 | llama-vulkan-radv | 1024 | 621.813 |
| 20 | bartwoski_Qwen3-Coder-Next-IQ4_XS.gguf | llama-vulkan-radv | Power Saver | 512 | 41.375 | llama-vulkan-radv | 1024 | 615.839 |
| 21 | Qwen3-Coder-Next-UD-IQ1_M.gguf | llama-vulkan-radv | Power Saver | 512 | 48.834 | llama-vulkan-radv | 1024 | 472.070 |
| 22 | kldzj_gpt-oss-120b-heretic-v2-MXFP4_MOE-00001-of-00002.gguf | llama-rocm7-nightlies | Power Saver | 512 | 40.004 | llama-vulkan-radv | 1024 | 432.180 |
| 23 | bartwoski_Qwen3.5-122B-A10B-IQ3_XXS-00001-of-00002.gguf | llama-vulkan-radv | Power Saver | 512 | 23.206 | llama-rocm7-nightlies | 1024 | 234.785 |
| 24 | unsloth-Qwen3.5-122B-A10B-UD-IQ3_XXS.gguf | llama-vulkan-radv | Power Saver | 512 | 20.771 | llama-rocm7-nightlies | 1024 | 194.398 |
| 25 | Qwen3.5-122B-A10B-Uncensored-HauhauCS-Aggressive-IQ3_XXS.gguf | llama-vulkan-radv | Power Saver | 512 | 27.562 | llama-rocm7-nightlies | 1024 | 186.736 |
Here is more detailed tables with exact context length for each run
https://pastebin.com/UU3rFKNA
