This review is built from ASUS and AMD specifications, retail pricing and third-party testing and estimates rather than hands-on use. Figures are attributed in the text. Some speed figures below are modelled estimates, and they are labelled as such.
The ASUS ROG Flow Z13 is the strangest local AI machine you can buy: a 13-inch gaming tablet with a detachable keyboard, built around AMD’s Ryzen AI Max+ 395 and available with up to 128GB of unified memory. That memory lets it load language models that no laptop graphics card can hold.
The honest answer is that the ROG Flow Z13 is a brilliant portable gaming and creative machine that also happens to run large models, rather than a dedicated AI tool. It shares Strix Halo’s 256 GB/s bandwidth, so it runs mixture-of-experts models well and dense 70B models slowly. A tablet’s cooling also limits how long it holds peak speed. In late 2026 it faces new competition from RTX Spark laptops and Apple’s M5 Max, and for many AI buyers one of those is the better fit. If you want one device for gaming, travel and occasional big-model work, it is still unique.
Table of Contents
- What the ROG Flow Z13 Offers
- ROG Flow Z13 Configurations and Price
- ROG Flow Z13 LLM Performance
- Why MoE Models Suit It
- ROG Flow Z13 Thermals and Sustained Load
- AMD’s Own Benchmarks, in Context
- ROG Flow Z13 vs Surface Laptop Ultra
- ROG Flow Z13 vs MacBook Pro M5 Max
- ROG Flow Z13 Gaming and Creative Work
- Running Models on Windows and Linux
- ROG Flow Z13 Battery and Portability
- Best Models for 64GB and 128GB
- Is It Still Worth Buying in Late 2026?
- Who Should Buy What
- FAQ
- Setting Up the ROG Flow Z13
- Verdict
- Sources and Further Reading

What the ROG Flow Z13 Offers
ASUS launched the 2025 Flow Z13 as the first device on AMD’s Strix Halo platform, and it remains the only tablet built around it. NoobFeed described it as the first of a new class of performance-focused small computers.
| Specification | ROG Flow Z13 (2025, top configuration) |
|---|---|
| Processor | AMD Ryzen AI Max+ 395, 16 Zen 5 cores |
| Graphics | Radeon 8060S, 40 compute units |
| Memory | Up to 128GB LPDDR5X-8000, unified |
| Memory bandwidth | 256 GB/s |
| Form factor | 13.4-inch tablet with detachable keyboard |
The unified memory is what makes it relevant for AI. NoobFeed explained the trade: you lose the split where desktop GPUs have their own VRAM, and in return you can configure up to 128GB that tasks such as local LLMs can use fully. Engineered’s summary of a review made the same point, noting that allocations far beyond gaming needs are aimed at LLM and content work.
ROG Flow Z13 Configurations and Price
The Flow Z13 came in several memory sizes, with 32GB, 64GB and 128GB versions on the Ryzen AI Max+ 395. For local AI, only the 64GB and 128GB versions matter, and 128GB is the one that runs the large models people buy Strix Halo for.
LLMRun listed the 128GB model at a launch price of about $2,799. Availability of that configuration has been uneven. A Reddit user in the Flow Z13 community wrote that the 128GB version was hard to find where they lived and settled for 64GB. Prices for high-memory devices rose across the market in 2026, so check current listings and expect the 128GB model to cost more than it did at launch if you can find one new.
The memory is soldered. If you might want to run 70B-class or 120B-class models, buy 128GB.
ROG Flow Z13 LLM Performance
Published figures for this device are a mix of measurements on Strix Halo hardware and modelled estimates. LLMRun’s estimates for the 128GB model are a useful summary: it runs 14B models and 120B mixture-of-experts models well, while 27B to 31B dense models and 70B models fit but are slow, at about 10 and 4 tokens per second.
| Model class | Approximate speed | Source and type |
|---|---|---|
| gpt-oss-120b, Q4_K_M | About 41 tok/s | LLMRun estimate |
| 120B mixture-of-experts on Strix Halo | Past 50 tok/s | Spawning Pint, measured on EVO-X2 |
| 27B to 31B dense | About 10 tok/s | LLMRun estimate |
| 70B dense | About 4 tok/s | LLMRun estimate |
Level1Techs covered the 128GB tablet in a video, and commenters on that discussion made the key point clearly: Strix Halo has about the same bandwidth as an M4 Pro but up to 128GB like an M4 Max, and bandwidth limits generation speed. Alex Ziskind’s review showed the tablet running a 110-billion-parameter Qwen model, which is the kind of thing no other tablet can do.
Why MoE Models Suit It
Mixture-of-experts models are the reason Strix Halo works for local AI. A 120B MoE model may activate only around 5B parameters per token, so the chip reads a small fraction of the weights each step. With 256 GB/s, that translates into comfortable reading speed despite the model’s size.
A Reddit commenter made this point early in the device’s life, noting that MoE models like Llama 4 Scout, which activates 17B of its 109B parameters, should run decently on Strix Halo. Newer models such as gpt-oss-120b and Qwen’s A3B line make the argument even stronger.
Dense models are the opposite. A 70B dense model reads all of its weights every token, and at around 40GB for a 4-bit version, 256 GB/s caps it at only a few tokens per second. If your work depends on dense 70B models, this is not the machine for you.
ROG Flow Z13 Thermals and Sustained Load
A tablet has much less room for cooling than a mini PC or a desktop. Spawning Pint’s Strix Halo review found that the ROG Flow Z13 in performance mode matches the desktop-class machines on shorter runs but throttles earlier under sustained load.
For chat and short coding questions, that rarely matters: each response is a short burst of work. For long agent sessions, batch processing or summarising a stack of documents, expect speeds to drop after a while compared with an EVO-X2 or Framework Desktop. Plug in, use the highest performance profile, and give the tablet airflow, ideally propped up on its kickstand rather than lying on a soft surface.
AMD’s Own Benchmarks, in Context
AMD used the Flow Z13 for its Strix Halo launch benchmarks, run in LM Studio on a 64GB unit. It claimed up to 2.2 times the token throughput of Intel’s Arc 140V, and up to 12.2 times the speed of an Intel Core Ultra 258V laptop on 14B models.
Those comparisons were chosen to flatter AMD. As one Framework community poster put it, AMD’s slides focused on models under 16GB, which is beside the point for anyone buying a 128GB Strix Halo machine; small models run much faster on a discrete graphics card. The useful takeaway is narrower: against other thin-and-light laptops without discrete GPUs, Strix Halo is in a different league.
ROG Flow Z13 vs Surface Laptop Ultra
The newest rival is Microsoft’s Surface Laptop Ultra with NVIDIA’s RTX Spark chip, which arrived in October 2026. It offers up to 128GB of unified memory with CUDA and around 300 GB/s of bandwidth, a little more than Strix Halo.
On paper the Surface should generate text slightly faster on large models and has the CUDA ecosystem. The Flow Z13 has x86 compatibility, mature Windows software, a tablet form factor and strong gaming. The Surface is a laptop running Windows on Arm, with some software gaps and no native Linux yet. Our Surface Laptop Ultra review covers it in detail.
ROG Flow Z13 vs MacBook Pro M5 Max
Apple’s MacBook Pro M5 Max with 128GB has up to 614 GB/s, more than twice Strix Halo’s bandwidth, so it generates text much faster on the same large models. It also has MLX, which can add large speed gains on Apple silicon.
It costs a lot more. The 128GB M5 Max starts several thousand dollars above the Flow Z13’s launch price. If large-model speed is your main reason for buying, the Mac is the better machine. If you want gaming, Windows, a tablet and good-enough AI, the Flow Z13 is the better value. Our MacBook Pro M5 Max review explains the Mac side.
ROG Flow Z13 Gaming and Creative Work
This is where the device earns its price for most buyers. NoobFeed measured around 120 frames per second in GTA 5 Online Enhanced at 2560×1600 on very high settings, staying above 100 with ray tracing. Tech Notice compared Strix Halo’s integrated graphics to a laptop RTX 4070.
For CPU work, NoobFeed placed Strix Halo between Apple’s M4 Pro and M4 Max in Cinebench. Video editing, photo work and code compilation all run well. As a single device for travel that games, edits and runs models, nothing else matches it.
Running Models on Windows and Linux
On Windows, LM Studio and Ollama run through Vulkan with no special setup, and AMD’s own benchmarks used LM Studio. Set the GPU memory allocation in AMD’s software so the integrated GPU can use most of the 128GB.
Linux works too. A Flow Z13 owner shared a tutorial for running Ollama and a web interface on Linux with the 128GB model, and reported that a coding model felt only a little slower than a cloud coding assistant. Recent kernels and drivers have improved Strix Halo support considerably.
As on all AMD platforms, try both the Vulkan and ROCm backends in llama.cpp if you use it directly, and keep the faster one.
ROG Flow Z13 Battery and Portability
A 13-inch tablet with a detachable keyboard is far easier to carry than a 16-inch workstation laptop. That is the Flow Z13’s biggest practical advantage over most AI-capable machines.
Battery life under AI load is short, as it is on every machine running large models. Sustained generation keeps the GPU and memory busy, and the tablet will also throttle sooner on battery. Treat unplugged inference as a convenience for a quick question, and do heavy work on mains power.
The tablet also gets warm under load, which matters if you hold it in your hands. For extended model work, use it docked with the keyboard on a desk.
Best Models for 64GB and 128GB
Choosing the right models matters more on this device than on a desktop GPU, because the wrong choice turns a fast machine into a slow one.
On the 128GB model, start with large mixture-of-experts releases. gpt-oss-120b is the obvious first test, estimated at around 41 tokens per second. Qwen’s larger A3B and A10B models and Gemma 4’s MoE variants are other good options. These give you quality close to big dense models at speeds that feel interactive.
On the 64GB model, aim lower. A 30B-class MoE model at 4-bit fits with plenty of room for context, and 14B dense models at higher precision run quickly. Dense 27B to 32B models fit but run slowly, around 10 tokens per second on Strix Halo, which is usable for careful work but tiring for chat.
On either version, avoid running several large models at once. Each one claims part of the shared memory, and Windows and your other apps need their share too. For a model-by-model list, see our best local LLMs guide.
Is It Still Worth Buying in Late 2026?
When the ROG Flow Z13 launched in 2025, it was the only portable way to get 128GB of GPU-accessible memory on Windows. That is no longer true. RTX Spark laptops now offer similar capacity with CUDA, and Apple’s M5 Max offers far more bandwidth.
It still has a clear place. It is the only one of these that is a tablet, the only one with full x86 Windows plus strong gaming, and often cheaper than the Mac with the same memory. A newer Strix Halo successor may appear, but the current chip remains the platform’s only top-end option today.
The decision comes down to how you will use it. Buy it as a gaming and creative device that also runs big models, and it is excellent value. Buy it as a dedicated AI machine, and a newer laptop or a desktop Strix Halo box will usually serve you better.
Who Should Buy What
| If you are | Start with | Why | Link |
|---|---|---|---|
| Wanting one portable device for gaming, creative work and big models | ASUS ROG Flow Z13, 128GB | Unique tablet with 128GB | Check price |
| Wanting the fastest large-model laptop | MacBook Pro M5 Max | More than twice the bandwidth | Check price |
| Needing CUDA in a portable machine | Surface Laptop Ultra | RTX Spark, up to 128GB | Check price |
| Not needing portability | GMKtec EVO-X2 | Same chip, better cooling, lower price | Check price |
Frequently Asked Questions About the ROG Flow Z13
Is the ROG Flow Z13 good for local LLMs?
Yes, for large mixture-of-experts models with the 128GB version. Dense 70B models fit but run at only a few tokens per second.
How much did it cost?
The 128GB model launched at about $2,799. Availability has been uneven and prices for high-memory devices rose in 2026.
Does it throttle?
Under sustained load, yes, earlier than desktop Strix Halo machines. Short tasks run at full speed.
64GB or 128GB?
128GB if you want 70B or 120B-class models. 64GB handles 27B to 32B models and smaller MoE models.
Is it better than the Surface Laptop Ultra?
For gaming, x86 software and tablet use, yes. For CUDA and slightly higher bandwidth, the Surface has the edge.
Can it run Linux?
Yes. Owners have run Ollama on Linux, and recent kernels support Strix Halo well.
Setting Up the ROG Flow Z13
Update the BIOS, drivers and ASUS’s Armoury Crate software first. Then set the GPU memory allocation high for AI work and choose the performance power profile when plugged in.
Install LM Studio or Ollama and start with a mixture-of-experts model such as gpt-oss-120b on the 128GB model, or a 30B-class MoE model on 64GB. Keep context length reasonable to save memory and speed.
For long jobs, keep the tablet upright on its kickstand on a hard surface so the fans can breathe.
Verdict
Buy the 128GB ASUS ROG Flow Z13 if you want one portable device that games, handles creative work and runs large mixture-of-experts models locally. Nothing else combines those in a tablet.
The honest caveat is that it is not the best pure AI machine. Bandwidth limits dense models, a tablet throttles earlier under long loads, many speed figures are estimates, and newer RTX Spark laptops and the M5 Max MacBook Pro offer more for AI-first buyers.
Next step: decide whether gaming and portability matter as much as AI. If they do, find a 128GB ROG Flow Z13 at a fair price. If AI comes first, compare the MacBook Pro M5 Max and the Surface Laptop Ultra before deciding.
Sources and Further Reading
- LLMRun: Flow Z13 128GB, for estimated model speeds and pricing.
- Spawning Pint: Strix Halo review, for measured speeds and the throttling finding.
- AMD: Ryzen AI Max+ 395 benchmarks, for the vendor test setup.
- NoobFeed: performance and productivity review, for gaming and CPU results.
- Framework community: Strix Halo expectations, for the bandwidth explanation.
On this site: compare portable options in our Surface Laptop Ultra and MacBook Pro M5 Max reviews, see desktop Strix Halo in the MS-S1 Max review, and pick models from best local LLMs.




