There's a particular kind of cognitive dissonance that comes with unboxing a machine sold on its ability to play Cyberpunk 2077 and Triple A games portably and immediately installing LM Studio on it. The ASUS ROG Flow Z13 (the 2025 model with the AMD AI MAX+ 395 and 128GB of unified DDR5 memory) is, on paper, a portable gaming device. In practice, it turned out to be the most capable local AI inference machine I've ever used, and a genuinely strong developer workstation in a form factor I didn't think was possible.
These are my field notes after putting it through its paces for AI workflows, web development, and daily school and freelance work.
Why This Machine
I needed a portable machine that could carry a full workload: freelance web development, school (final semester of a CS/IT degree), and serious local AI inference, all at once, without compromise.
That's a harder brief than it sounds. The unified memory story is what makes the Z13 unusual. The AI MAX+ 395 and 128GB of shared DDR5 is, at time of writing, one of the very few portable configurations that can hold and run large language models entirely on-device, without offloading to CPU or splitting across multiple machines. For a machine you can actually carry, that's a short list.
I was also coming from a Framework 13 evaluation phase. I spent time looking at it for its repairability story and Linux-first posture, but the memory ceiling wasn't there for the inference workloads I had in mind. The Z13 was the call.
The Hardware, Briefly
The AI MAX+ 395 is a peculiar chip. It's an AMD part that packs 16 CPU cores (Zen 5), 40 Compute Units of RDNA 3.5 GPU (branded as a Radeon 8060S in the Z13 config), and an NPU, all sharing a single 128GB DDR5 memory pool. There's no discrete GPU memory to speak of. Everything (CPU data, GPU framebuffer, model weights) lives in the same flat address space.
For most GPU workloads, that's a constraint. For local LLM inference, it's a superpower. A 70B model quantized to Q4 comes in around 40GB. On a conventional setup, you'd need 40GB of VRAM to hold it, which means a very expensive multi-GPU server rig. On the Z13, you just... load it. It fits in RAM. The GPU can see it directly.
Other notables:
- 13.5" 2560×1600 IPS, 180Hz, DCI-P3 100%: an absurdly good display for a machine this size
- M.2 2230 internal slot: proprietary-length, which I'll get to
- USB4 ×2, USB 3.2 ×1, UHS-II microSD: solid port selection for a thin machine
- Detachable keyboard: tablet form factor is the Z13's defining quirk
I added a Crucial P310 2TB NVMe in the only M.2 slot. One important callout: the Z13 only accepts M.2 2230 (30mm) drives, not the standard 2280. The P310 happens to come in 2230. If you're planning storage expansion, check your drive length before buying.
LM Studio and Local Inference: The Main Event
This is where the Z13 earns its keep.
I run LM Studio as my primary inference front-end, using the Vulkan backend on Windows. The VGM (Video GPU Memory) allocation needs to be set correctly; LM Studio and similar tools need to know how much of the unified pool to treat as GPU-accessible. On Windows with Vulkan, this is configurable and makes a real difference in whether a model runs on the GPU at all versus falling back to slower CPU inference.
Models I'm actively running:
- GPT-OSS 120B (Q4): the flagship of my stack; 120B at 4-bit quantization fits and runs at a pace I'd actually call interactive
- GPT-OSS 20B (Q4): faster for quick tasks where the bigger model is overkill
- Qwen3.5 35B A3B (Q4): a Mixture-of-Experts architecture that punches well above its active parameter count; great multilingual capability
One thing I landed on quickly: anything above Q4 quantization is too slow for my taste. I'd rather have a larger model at Q4 than a smaller one at Q8, and I'd rather have fast, usable output than ticker-tape generation from a higher-fidelity quant. The Z13 makes that trade-off sensible; there's enough memory to hold a genuinely large model at Q4 without compromise.
Practical throughputs (rough, varies by context length):
- 20B Q4: fast, responsive, good for rapid iteration
- 35B Q4 (MoE): excellent quality-to-speed ratio
- 120B Q4: noticeably slower but interactive; earns its keep on complex tasks
For a portable machine with no external power requirement beyond a USB-C charger, this is remarkable.
The Developer Workflow
Local AI inference is one thing. The rest of the stack matters too.
Web development: I work primarily in PHP (Laravel), Python, and JavaScript/TypeScript. The Z13 handles local dev environments without complaint; Docker containers, MariaDB, Vite dev servers, all running concurrently without the machine breaking a sweat. I've had LM Studio running a 35B model in the background while actively coding in VS Code and it stayed responsive.
Laravel specifically: I'm running taucetiiv.net, a Marathon fansite built in Laravel with Livewire and Alpine.js. Development on the Z13 is smooth. The 16-core Zen 5 CPU handles compilation and build steps quickly enough that I don't miss desktop performance.
University coursework: This semester covers C, Linux, and Data Science. Having to context-switch between C and PHP in the same day is where the Z13's snappiness earns its keep; nothing feels sluggish, and Rust tooling (which I've been picking up on the side) compiles fast enough that I'm not watching a progress bar. The data science workloads that would otherwise need a cloud notebook can just run locally.
The CalDigit TS4 dock: I use this as my desk anchor; it connects to both USB4 ports and provides 2.5G ethernet, multiple display outputs, and charging. Stability on Windows has required some attention; the dock occasionally needs reconnection after resume from sleep. It's a known Windows interaction, not a TS4 issue per se. Still a net positive to have a proper docking setup.
The Tablet Form Factor: Weird, But It Works
The Z13 is genuinely a tablet. The keyboard detaches completely. The display can be rotated to portrait mode. There's a kickstand.
For a developer, this is mostly a quirk rather than a feature. The detachable keyboard is fine to type on, though it's not my favorite. I'm a fast typist (sitting around 220 WPM on a full-size board) and the Z13's keyboard cuts into that. Not dramatically; it's better than most laptop keyboards, but it's the one area where the form factor makes a real trade-off.
Where the tablet mode actually earns its place: reading long documentation, reference-checking while coding on an external monitor, and the occasional use case of having the display propped at an unusual angle on a desk. The 2560×1600 display is sharp enough that it's pleasant to read at arm's length.
It also makes the machine easier to carry. 13.5" is already portable; without the keyboard it genuinely fits anywhere.
Thermal and Power Behavior
Thermals on the Z13 are better than I expected for a chip this capable in this chassis; it doesn't run hot in the way that gaming laptops often do. The vapor chamber is doing its job. What I did notice is that the fans can get louder than I'd like under sustained inference load, which is a minor annoyance when working in a quiet environment.
I run exclusively in Turbo mode when plugged in. No reason not to; the chip has headroom and the dock provides plenty of power. On battery, I've been using Performance mode by default, and battery life under that profile is predictably short for gaming. I haven't yet tested lighter profiles on battery, so I suspect Silent or Standard mode would extend that meaningfully for lighter tasks. That's an experiment for later.
The power management story on Windows is still evolving. AMD has published updated drivers and ASUS pushes BIOS updates regularly. It's worth keeping both current; the situation improves over time.
Who Should Consider This Machine
Yes, if:
- You want to run large models locally on a portable machine
- You work across coding, AI, and general productivity
- The tablet form factor doesn't put you off
- You're comfortable tuning drivers and power profiles
No, if:
- You need maximum single-threaded CPU speed
- You need > 2230 M.2 drives (the slot is 2230 only)
- You want a conventional clamshell laptop experience
- Your workflow benefits more from raw GPU bandwidth than memory capacity (video transcoding, certain simulation workloads)
Closing Thoughts
The ASUS ROG Flow Z13 with the AI MAX+ 395 occupies a strange and genuinely new niche: a portable machine that can run frontier-scale local AI models without compromise, while also handling a full developer workload across multiple languages and environments. That the same chip drives an excellent display and fits in a bag without drama makes it more than a one-trick pony.
It's not perfect. The form factor asks for some buy-in. The fans have opinions under heavy load. But for the specific combination of local AI inference + polyglot development + portability, it's the most capable thing I've found at this price point.
Simon is a freelance full-stack developer and final-semester CS/IT student. He writes about hardware, AI tooling, and web development.