LLM on the go - Testing 25 Model + 150 benchmarks for Asus ProArt Px13 - StrixHalo laptop

Posted by Willing-Toe1942@reddit | LocalLLaMA | View on Reddit | 7 comments

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

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

[-]

Appropriate_Ad1792@reddit

I have the same, biggest downside is the 2230 ssd, that is pcie 4.0. Now maximum ssd's are 2tb... there is a micron 4tb pcie 5.0 that will be released but it will be expensive and limited to 7500 mbs by the controller. Also, only now I have seen that 2230 ssd's don't hqve ddram cache, that makes them 2x slower than normal ones... Adding that screen has 60hz only, and that the apu is limited to 75w... (in windows you can overclock to 85w with their tool and set up max 115w for secconds... It does not have vapor chamber... but yes, for 13" is the best laptop in this size.

I would have got "HP zBook Ultra G1A Ryzen AI Max+ 395 , 128GB RAM, 2TB SSD" that has all those fixed but at 14" but it was like 2000 usd more expensive 😅

lpdand@reddit

I wish there were more options on the Ryzen AI Max+ 395 platform.

I can't thank you enough for this.
I'm looking for a new daily driver laptop for coding + ability to run local LLMs and this PX13 with 128GB seems to be a very good option.

Willing-Toe1942@reddit (OP)

you will love it I use Qwen3.6 unsloth on it and oh-god it's a beast with pi-coding agent. I can easily process 12 M token nearly each hour (thanks to prefix Caching)

That's awesome. Thanks for the info!

here is yesterday pi-coding consumption summary to give you an idea about how much work you can handle it daily

king_of_jupyter@reddit

Damn nice