It's one writer's hands-on account. It isn't a controlled benchmark, and I haven't reproduced it. Here is what he reports, what the spec sheets and other reviews add, and what to check before you buy.
What the author actually claims
Campanale's verdict is that the Zenbook is not the faster of the two machines. In his testing, though, it loads models the 32GB RTX Spark laptop won't. He says the RTX Spark ran models quickly once they loaded, but struggled to load moderately sized ones. The models he wanted were too large to use with agent platforms like Hermes, which need a bigger context window.
The Snapdragon machine, by his account, ran everything he threw at it. That included Qwen 3.8 27B, which he calls one of his favorite local models right now. He also says the Zenbook weighs about half what the OmniBook does, has better battery life, and has a screen that's essentially as good.
That is a memory-headroom result. A 32GB machine must fit the model weights, the context cache and the operating system into one pool. A 48GB machine has 16GB more room. It would be a mistake to read this as "Snapdragon beats NVIDIA."
Don't generalize the RTX Spark result
The comparison is against one RTX Spark configuration. NVIDIA's launch announcement describes the platform as offering up to 128GB of unified memory and up to 1 petaflop of FP4 AI performance. Third-party listings say the 48GB X2 Elite Extreme can run Qwen3.8 27B, which fits the author's experience. But a higher-memory RTX Spark SKU would change the picture, so the real takeaway is "buy enough RAM," not "avoid RTX Spark."
NVIDIA says RTX Spark laptops are available from ASUS, Dell, HP, Lenovo, Microsoft and MSI, starting at $2,599. NVIDIA also names Hermes and OpenClaw among the agents with day-one support. The lineup includes the HP OmniBook Ultra, which is the family the author tested. Prices and configurations vary by region.
The Zenbook A16: which one is the one he's talking about?
Asus lists the Zenbook A16 (UX3607) in several configurations, and they aren't interchangeable:
- Top tier: Snapdragon X2 Elite Extreme X2E-96-100 or X2E-94-100 with 48GB of on-package LPDDR5X. Asus lists 228GB/s memory bandwidth, a Hexagon NPU of up to 80 TOPS, and a 3K (2880 x 1800) 120Hz OLED.
- Entry tier: an 8-core Snapdragon X1-26-100 with 32GB of memory, a 45-TOPS NPU and a 60Hz WUXGA OLED.
- Weights: Asus lists 1.30 kg (2.87 lb), 1.20 kg (2.65 lb) and 1.10 kg (2.43 lb) across the variants. The author's 2.65 lb figure matches one of those entries.
So "Zenbook A16" doesn't guarantee the Extreme chip or 48GB. If you're shopping for local AI, check the exact model number.
Price: a moving target
The author says the Zenbook costs about half as much as his OmniBook, "especially when sales are going on," and says it's under $2,000. His page lists $1,900 at Best Buy and $2,200 at Asus. Those are the figures his page showed, and they will change.
Tom's Hardware tested an X2E-94-100 configuration with 48GB and 1TB of storage at $1,699.99. It said the pricing had been revised upward on launch day from a promised $1,599.99. Treat all of these as snapshots, not current pricing. Check the retailer before you decide.
Independent reviews give a more mixed picture
The reviews the author's account sits beside are less glowing about the rest of the laptop.
- Tom's Hardware called the chip speedy but the package so-so. It criticized a chassis that flexes and a shorter battery life than competitors, and said Windows on Arm still has some hiccups.
- TechRadar found general performance respectable but graphics disappointing. The A16 failed Puget Bench's Adobe Premiere benchmark and wasn't playable in Cyberpunk 2077 at its Ray Tracing: Low preset. It was also impressed by battery life, at 24 hours 5 minutes in its movie-loop test.
Battery figures from different test methodologies don't compare directly. They also aren't a head-to-head against the OmniBook, so the author's battery claim is neither confirmed nor refuted. The pattern across all the reviews is that the A16 is a light, efficient general-purpose machine with strong memory capacity and integrated graphics, not a graphics workstation.
"80 TOPS" doesn't mean what you'd hope
Marketing leans on NPU TOPS, but they don't map neatly to local LLM speed. One third-party tool listing, llmrun.dev, warns that the 80-TOPS NPU is not used by llama.cpp or Ollama, and that local LLM speed is bound by roughly 228GB/s of memory bandwidth, running on the CPU or Adreno GPU. That is a community-style source rather than an official one, so treat it as a lead. It matches what independent users have reported on the earlier Snapdragon X Elite generation. One user of a Surface Laptop 7 wrote that Ollama on Windows ARM uses llama.cpp, which has no NPU backend.
NPU-first runtimes do exist. One open-source project describes an NPU-first local LLM runtime for Snapdragon X Elite on Windows on ARM, and another write-up walks through using Microsoft Foundry Local, installed through WinGet. These tend to target smaller models. If your plan depends on the NPU, verify that your chosen software actually uses it.
That is also why the author's result is plausible. On this hardware, a 27B model is mostly a question of whether it fits in memory and how fast the memory feeds it. Those are facts about capacity and bandwidth, not about the TOPS number on the box.
A practical buying checklist
- Start with the model you want to run. Work out its size at your chosen quantization, then add room for the context window. Agent frameworks use larger contexts, which is the problem the author hit.
- Buy memory first, branding second. A cheaper machine with 48GB can beat a pricier one with 32GB for loading big models. On the other hand, a faster platform with more memory would win if you can afford it.
- Confirm the exact SKU. On the Zenbook A16, the memory, chip, display and weight all vary.
- Check software support on Windows on Arm. Qualcomm's chips are ARM64. Confirm that the runtime you want (Ollama, LM Studio, Foundry Local or an agent framework) supports ARM64 and what hardware it targets.
- Know your other workloads. If you also edit video or play games, the A16's integrated graphics could limit you, per TechRadar.
- Treat sale prices as temporary. Compare at checkout.
Bottom line
Campanale's argument is credible within its limits. For a 32GB RTX Spark machine, loading larger models and big context windows is a real constraint. A 48GB Snapdragon laptop that costs less and weighs less is a sensible option for local AI. It doesn't show that Snapdragon is faster, that it's better than every RTX Spark or Strix Halo machine, or that its NPU is doing the work. If your goal is to run bigger models on the go without spending $2,600 and up, the Zenbook A16 deserves a look. Just verify the configuration, the price and the software support before you buy.