ROG Flow Z13 2025 128GB - Qwen3.8 Flash-Next benchmarks with Halogen, Gufo and Rulith (60W vs 93W, up to 200K context)
Been playing around with the Strix Halo LLM projects that have been popping up lately, so I spent a few hours seeing how far I could push Qwen3.8 Flash-Next on my 128GB 2025 Flow Z13.
Tested Halogen, Gufo and Rulith at both 60W and 93W.
60W is what I actually consider usable day to day on the Z13. 93W is more of a "let's see what it can do" setting.
One thing before the numbers: this is not a strict apples-to-apples backend comparison. The model formats and MTP setups are different between them. I'm mostly comparing the setups I'd realistically use with each project right now, not trying to prove backend X is universally faster than backend Y.
Test setup
- ROG Flow Z13 2025 — Ryzen AI MAX+ 395, 128GB
- CachyOS KDE — kernel 7.2.8-2-cachyos, UMA 512MB, GTT 124GB, IOMMU off
- Halogen 0.16.0 — Qwen3.8 Flash-Next v2 checkpoint, MTP depth 2
- Gufo 0.5.0 — Qwen3.8 Flash-Next UD-Q4_K_XL, MTP Q8_0 draft 7, 262K context / 1 session
- Windows 11 25H2 + Rulith 0.4.1 — 32GB RAM / 96GB VRAM, UD-IQ4_XS, MTP draft 3, 262K context
Halogen
Halogen was easily the fastest one for prompt processing/prefill in my testing.
93W
At 64K I got about 1722 tok/s PP.
At 200K it was still doing 1694 tok/s PP and 41.16 tok/s decode.
What surprised me more than the peak number was how little PP dropped with context size. Going from 64K to 200K was basically:
1722 -> 1694 tok/s
So for long context, this setup held up really well.
60W
At 60W I got about 1573 PP / 45 tok/s decode at 64K, and 1509 PP / 35.4 tok/s at 200K.
This is probably the setting I'd actually use.
The fans calm down quite a bit and thermals are obviously easier to deal with, while prompt processing is still over 1500 tok/s at 200K.
60W feels like the sweet spot on this machine.
Gufo
Gufo was actually the one I was most curious about before doing this.
93W
Around 1398 PP at 64K, and 1343 PP / 33.6 tok/s decode at 200K.
Still fast, but there's a pretty noticeable gap to Halogen with the configs I tested.
60W
Around 1238 PP at 64K, and 1156 PP / 33.0 tok/s at 200K.
Gufo seemed to care more about the power limit than the other two, at least for PP.
At 200K it went from roughly 1156 PP at 60W to 1343 at 93W, so about a 16% increase.
The project is moving pretty quickly though, so I'm definitely going to keep testing it. Right now, on my Z13 and with these settings, it isn't catching Halogen yet.
Rulith(old strix llama) / Windows
Rulith was interesting for a different reason.
At 93W I got about 1221 PP at 64K and 1170 PP / 32.55 tok/s at 200K.
So PP is clearly behind the Linux setups here.
But short-context decode is actually very fast. At 1K context I saw 54.7 tok/s.
The downside is that decode falls off pretty hard as context grows, ending up at about 32.55 tok/s at 200K.
So the pattern I saw was basically:
short context: Rulith is very fast
long context: Halogen holds its speed much better
Rulith at 60W
This is actually closest to how I normally use the Z13.
About 1138 PP at 64K, and 1070 PP / 31.29 tok/s at 200K.
I already have a dual CMP 170HX box running 24/7 as my headless LLM server, so the Z13 is more of a main PC for me.
I do have CachyOS set up as a dual boot, but switching back and forth is a little annoying because I have to change the memory allocation in BIOS depending on what I'm booting.
Because of that I end up leaving it in Windows most of the time, and Rulith is honestly pretty convenient for that.
All of them together
For PP, Halogen is pretty clearly ahead with the setups I tested.
At 200K:
Halogen 93W: 1694 PP
Halogen 60W: 1509 PP
Gufo 93W: 1343 PP
Gufo 60W: 1156 PP
Rulith 93W: 1170 PP
Rulith 60W: 1070 PP
Decode is more interesting.
Rulith starts out very fast at low context, but drops quite a bit as the context fills up.
Halogen is much flatter across the whole range.
So for normal short chats, Rulith can actually feel really good. For coding/agent stuff where the context keeps growing, I'd rather have Halogen.
Overall
For my current setups:
Halogen: best PP and best long-context behavior.
Rulith: slower PP, but very convenient if you actually want to use the Z13 as a Windows PC, and short-context decode is surprisingly good.
Gufo: not as fast as Halogen yet on my machine, but probably the one I'll keep watching because development seems pretty active.
Strix Halo 128GB is still kind of a weird niche, but I think that's also what makes it interesting.
If you want enough VRAM for models like Flash-Next on conventional consumer GPUs, the cost gets stupid pretty quickly. There are obviously more VRAM-efficient backends and free-token/offload tricks showing up now, so it's not a simple price comparison anymore.
Still, being able to run Qwen3.8 Flash-Next with a 200K context on what is basically a 13-inch tablet/laptop is pretty wild.
I paid about 3.2M KRW used($2400) for this Z13. New Z13s and the ProArt/PX13-type Strix Halo machines also occasionally show up below 4M KRW($3000) here, so for someone specifically interested in local LLMs, I think they're in a pretty unique spot.
Of course the actual problem is that I bought this thing for local LLMs and somehow ended up using it as a gaming machine more often.
