Wire · AZ Weekly PostZDNET’s key takeaways Lenovo’s new RTX Spark laptop handled an outrageous local AI coding test. The gorgeous OLED screen and generous ports add everyday appeal. Heavy workloads expose memory limits and a loud fan.
Let’s get the regular ol’ laptop review stuff out of the way: this is probably the nicest $4,000 laptop I’ve ever reviewed. But this is no regular laptop. Oh, no.
It’s powered by Nvidia’s RTX Spark processor, which changes the PC game entirely. The Lenovo Yoga 9n 2-in-1 16-inch laptop is one of a new crop of machines based on Nvidia’s new hotness. The machine I’m testing includes a 20-core NX1 processor with 64GB RAM.
What makes this machine special compared to other chips is Nvidia’s powerful Blackwell RTX 6144-core GPU with CUDA. CUDA (Compute Unified Device Architecture) is Nvidia’s system that allows developers to use GPUs for computing tasks. More from ZDNET Also: Microsoft’s new Surface Laptop Ultra finally has a starting price (you should sit down) Benchmarking the AI Let’s start off with some basic benchmarks.
For this review, I decided to pit the $4,000 RTX Spark-based Lenovo laptop against my daily driver, a 2025 Mac Studio. This comparison is not entirely fair. My Mac Studio cost me $5,300 in April of 2025.
This previous-generation machine is equipped with an Apple M4 Max chip, a 16‐core CPU, a 40‐core GPU, 128GB of unified memory, and an 8TB SSD. By comparison, the 20-core Yoga has only 64GB of RAM and a 1TB SSD. Also: I thought I had the new Siri AI in iOS 27 – until I found it hidden in my settings It’s the RAM that will come to haunt these tests.
Unfortunately, the Yoga can’t be equipped with more than 64GB of RAM. On the other hand, this is a laptop. For something portable, it’s pretty amazing.
I started with Geekbench. As you can see, the Mac delivers about 1.3 times the CPU performance vs. the Yoga in both tests. Geekbench comparison.
I also attempted to run Geekbench’s AI benchmark. It failed because it insisted on measuring the Yoga based on Intel emulation, not native Arm performance. So I switched to Llama-bench , a tool that measures actual language model performance (both prompt processing and token generation) using the same models and workloads across platforms.
This gives a very solid apples-to-apples (or, well, Apples-to-Yogas) metric for how one machine performs compared to the other. Also: How I used ChatGPT to talk me through fixing a corrupted gaming PC – literally The first test I ran was prompt processing running entirely on-machine. This measures how quickly the AI processes and understands the prompts.
Faster performance is helpful for really long prompts. Here, the Lenovo was about 3x faster than the Mac. That’s pretty astonishing, but it doesn’t tell the whole story.
Prompt processing benchmark. That’s because the token-processing metric tells us how fast the AI actually performs AI work. For example, how quickly can it generate code, or do some agentic workload using entirely local on-machine AI?
In this test, the Mac pulled out ahead. You can see evidence of the memory squeeze: the Mac was only about 1.5x faster with an 8-billion-parameter model (Qwen3-8B-Q4_K_M), but 2.5x faster with a 32-billion-parameter model (Qwen3-32B-Q4_K_M). Clearly, more RAM helps the model fit better.
This became apparent not only here, but in actual workloads. Token processing benchmark. Even so, I’m very impressed that a laptop can hold its own against a heavily equipped Mac Studio.
Can it code? Oh yeah. It can code.
Knocked my socks off, in fact. More on that in the next section. To get started, I set up an entirely local coding environment on the Yoga consisting of: Coding agent: OpenCode v2.1.4 AI coding model: Qwen3-Coder-Next 80B-A3B, Q4_K_M quantization Inference engine: llama.cpp with Nvidia CUDA acceleration Context window: 32,768 tokens I tried a bigger context window.
But even though the Yoga has 64GB, with Windows installed and all the overhead, the most I could get to work was the 32K token window. Keep in mind that I was also using an 80-billion-parameter Qwen model. OpenCode on the Lenovo.
There is no doubt this felt restrictive. I’ve been coding with Claude Code and OpenAI Codex, using their cloud-based AIs. I regularly chew through a 1-million-token context window a few times a night.
Here, on the Yoga, I was running 1/30th of that context window. This means that OpenCode could process only one or two actions before it had to record the full status of the build and reload the context. That was tedious to watch.
It also slowed down result processing considerably. A completely bonkers programming test I gave the Yoga a very unfair test. I copied all the source code from my Icon Gallery Mac App Store project and dropped it into a folder on the Windows machine.
Let me explain how nutball this is. I brought over Mac-only source code and a Mac-only framework. Nothing of that could possibly work on Windows.
Then, with no spec at all (I didn’t even give it any screenshots of what the Mac product looked like), I simply told OpenCode to make it work on Windows. It had to derive everything it knew from source code that doesn’t even belong on this architecture. Also: How I made a paid Mac app in 11 days with Claude – without writing a single line of code OpenCode decided to move the app from Swift to C#.
I thought that would be convenient for me, because I taught C# at UC Berkeley for more than a decade and I know the language quite well. It turns out that my skill set wasn’t even necessary because the AI coded everything all on its own. I didn’t write a single line of code or edit a single config file.
At first, OpenCode built using the legacy WPF framework because it provided a faster path for it to figure out how to parse the encrypted binary icon data and define core functionality. OpenCode and Qwen reporting on progress. Once that was stable, OpenCode migrated to WinUI 3 to incorporate modern deployment capabilities and align with Microsoft’s Windows UI strategy.
Within about four hours, it was able to render the icons and roughly imitate the original Icon Gallery window. Not a bad facsimile of the original project. Clearly, there’s a lot more work to be done to make a distributable Windows app, which I don’t plan to do.
Even so, you can see that the entirely local AI performed admirably, producing something that came recognizably close to the original app. The original app, running on a Mac. I was a bit frustrated with the constant context refreshes, but even with that inconvenience, I was blown away by how far an entirely local AI was able to get on a laptop.
Keep in mind that as recently as February, I tried using local AI coding, and it failed miserably. Also: I tried a Claude Code rival that’s local, open source, and completely free – how it went Clearly, the models, the frameworks, and the hardware have evolved quite a bit. I am blown away.
I was also impressed by the machine’s heat management. After running it for four hours of intense local AI coding, the Yoga laptop wasn’t hot to the touch. Most of the surface was barely warm.
The keyboard itself was slightly warmer, but nothing you could remotely consider hot. Try that with an Intel processor. The whole desk would probably burst into flames!
A somewhat unfair battery test So, then I decided to see how well the AI would perform on battery power. Lenovo doesn’t specify how long their battery is supposed to last. On the main laptop page , Lenovo says, “Supports Rapid Charge Express (15 minutes = three hours capacity).” But is that three hours, or is it more?
Is it still three hours if you’ve only charged it for exactly 15 minutes? I decided to stress the laptop and see how long it would last. I decided to do this by loading the Hermes agent, which you can actually download from the Microsoft App Store.
Hermes in the app store. Amazingly, Hermes configured almost completely out of the box, including the Qwen 27B model. I answered a few simple questions, and it just worked.
After all the time I spent setting up OpenClaw on a dedicated Mac, this level of setup was insanely great. I decided to turn Hermes to evil. I asked it to build an AI test case that would drive all the device’s GPUs and CPU cores to see how long the machine would run.
As it turned out, Hermes was able to run the AI to provide prompt parsing and responses. As we saw previously from the prompt benchmarks, this machine is really good at that. Unfortunately, as soon as I tried to put a real AI workload against the battery test, it failed hard.
So, yeah, that didn’t work. So I took a different approach and instructed Hermes to build a Python script to exercise the machine. It set up a test pushing raw CPU work consisting of 18 processes hammering SHA-256 hashing, calling OpenSSL code, plus one process churning through memory.
Let’s use some CPU. As you can see, it gave the machine a workout. I also set up a concurrent time logger, which wrote the time to a file every five minutes.
That kept going until the machine crashed from running out of power. After plugging it in and rebooting, I was able to gather some performance details. Also: Use ChatGPT Plus?
Switching to Gemini could cut your bill – by $144 a year for me The test started at 19:35 with 100% battery. It continued running until 20:55, a total of one hour and 20 minutes. Beyond the testing details, it was pretty easy to tell the machine was running at full load.
Its fan is seriously loud. When it started up, I could hear it in the next room. I have to give this machine credit.
It’s a tough little machine. It ran for almost 90 minutes on a load designed specifically to bring it to its knees. That’s impressive.
More about the hardware The 16-inch 2880 x 1800 OLED screen is simply gorgeous. The touchscreen is responsive, although I spent most of my time using the keyboard and trackpad. The trackpad is also very responsive and large.
Navigating was crisp and easy. The Yoga part of the laptop is Lenovo’s popular transforming capability, where the machine can be a regular laptop, be folded like an easel, or even be turned into a tablet. Living up to the Yoga name.
The machine has two USB-A ports (up to 10Gbps), one on each side. It also has two USB-C ports, also one on each side. The left side has a full-size HDMI port and a headphone jack, and the right side has a full-size SD card reader.
Networking is solid. The machine comes with Wi-Fi 7 and Bluetooth 5.4. I have pretty good Wi-Fi in most rooms of my house.
But to take pictures, I had to bring the laptop into my workshop, which is something of a Wi-Fi dead zone. I was easily able to connect to it from my desk, with absolutely no interruptions or glitches. Other laptops in that location have struggled, so points to Lenovo for good Wi-Fi.
Also: Buying your next Windows laptop? This Lenovo with a tandem OLED display is my top pick Compared to my 13-inch MacBook Air, which weighs 2.7 pounds, the 16-inch Yoga is heavy. But at 4.28 pounds, the Yoga is almost half a pound lighter than Apple’s 16-inch MacBook Pro, which weighs 4.7 pounds.
Keep in mind that this Yoga is a machine that can locally port a Mac app to a Windows app, all on its own without cloud access. When you factor that in, the extra weight is well worth the extra oomph. My only real criticism of the hardware is the silly placement of the pen.
The machine comes with a little removable magnetic pen holder. It looks like a charger, but it’s not, it’s just a cabin for the pen itself. You charge the pen by plugging a USB cable into the top of it, but it has to be outside the holder to charge.
This is the only place the pen seems to attach. The Yoga laptop is 17.5 mm thick. The pen is only 9.13 mm thick, which means there’s theoretically enough space inside the spine of the machine to hold the pen.
But the company instead chose to include a magnetic containment device that can easily be knocked off. Given that the pen, separately, costs around $100, it’s not something you want to take a chance on losing. Worse, it only attaches to the back of the laptop, which means it’s not accessible in either of the two configurations where users are most likely to want to use the pen.
You would think they’d include a nice slot for the pen to live inside. Final thoughts The Lenovo Yoga 9n 2-in-1 16-inch is a killer machine that can actually host and run local AI. But the max 64GB of memory is a definite limit to running heavy-duty AI work entirely locally.
Nvidia lists the max capability of the RTX Spark as 128GB. If Lenovo wants to be a major player in the local AI world, I would strongly recommend they add a 128GB model to their Yoga 9n offerings. Also: Could AI really destroy us all, or are humans still the bigger threat?
Even with 32GB or 64GB of RAM, you can get a lot done on this machine. I wouldn’t expect to do entirely local AI coding without some annoyances. The fact that it works, and that it was able to (somewhat) port an app from Mac to Windows, converting the language and framework along the way, is an astonishing testament to what is now possible — and what will be possible in the future — with this platform.
It stays cool under heavy load, it has a fantastic display, it’s fast and responsive. What’s not to like? If you’re a Windows user with the budget, and you’re fine with 64GB of RAM, I can easily recommend this machine.
You can follow my day-to-day project updates on social media. Be sure to subscribe to my weekly update newsletter , and follow me on Twitter/X at @DavidGewirtz , on Facebook at Facebook.com/DavidGewirtz , on Instagram at Instagram.com/DavidGewirtz , on Bluesky at @DavidGewirtz.com , and on YouTube at YouTube.com/DavidGewirtzTV .
Source: ZDNET
