The machine
The unit is a Lenovo Yoga 9n 2-in-1 16-inch. It has NVIDIA's 20-core RTX Spark chip, a 6,144-core Blackwell GPU, 64GB of unified memory and a 1TB SSD. Gewirtz puts the price at about $4,000. Unified memory means the CPU and GPU share one pool, which matters a lot for local AI.
Other coverage of this model gives context:
- How-To Geek reports the 2-in-1 tops out at 64GB of RAM and 2TB of storage.
- The sibling Yoga Pro 9n can go up to 128GB of RAM.
- The Gadgeteer lists the Yoga 9n 2-in-1 at $3,749.99. That is a different price from the roughly $4,000 Gewirtz cites, probably because configurations differ.
So the memory limit is specific to this model. It is not an RTX Spark limit. NVIDIA's own materials say the platform goes up to 128GB.
The coding test
Gewirtz set up a fully local stack:
- Agent: OpenCode v2.1.4
- Model: Qwen3-Coder-Next 80B-A3B at Q4_K_M quantization
- Inference: llama.cpp with CUDA acceleration
- Context window: 32,768 tokens
He says larger context windows would not work once Windows and its overhead were counted against the 64GB. He compares 32K tokens to the roughly 1-million-token windows he regularly uses with cloud tools. That is about one-thirtieth. In practice, OpenCode could handle only one or two actions before it had to save its state and reload context. He calls this tedious and says it slowed results.
The task was to port his Icon Gallery Mac App Store app to Windows. He copied over the Mac-only source and a Mac-only framework. He gave no spec and no screenshots, only an instruction to make it work on Windows.
- OpenCode chose to move the app from Swift to C#.
- It first built on WPF, which Gewirtz says gave a faster route to parsing the app's encrypted binary icon data.
- It then migrated to WinUI 3.
- After about four hours it could render the icons and roughly imitate the original window.
- Gewirtz says he wrote no code and edited no config files.
He is clear that this is not a finished app. He says much work remains before it could be distributed, and he doesn't plan to do it. I'd read it as a promising prototype. Nobody has checked it for correctness, security or edge cases.
Benchmarks against a bigger Mac
He compared the Yoga with his own 2025 Mac Studio. It has an M4 Max, 128GB of memory and an 8TB SSD, and cost $5,300. He calls the comparison unfair himself.
| Test | Reported result |
|---|---|
| Geekbench CPU | Mac about 1.3x faster in both tests |
| Geekbench AI | Failed on the Yoga, which it measured as Intel emulation rather than native Arm |
| Llama-bench prompt processing | Yoga about 3x faster |
| Token generation, Qwen3-8B Q4_K_M | Mac about 1.5x faster |
| Token generation, Qwen3-32B Q4_K_M | Mac about 2.5x faster |
Prompt processing is how fast the model reads your input. Token generation is how fast it writes output, which is what you wait on during code generation. Gewirtz blames the widening gap on memory pressure. That is plausible, but he did not isolate it as the cause. The article has no full tables or methodology, so treat the ratios as indicative.
Thermals, battery and noise
- Thermals: After four hours of intensive local AI coding, he says most of the surface was barely warm and the keyboard only slightly warmer.
- Corroboration: StorageReview found the Yoga's CPU peaked between 71 and 86 C under all-core loads that hold the HP OmniBook at 98 C. That broadly fits Gewirtz's cool-running impression, though StorageReview measured component temperatures and Gewirtz described how the surface felt.
- Battery stress test: He had the Hermes agent write a Python script with 18 processes doing SHA-256 hashing through OpenSSL, plus one process churning memory. A logger recorded the time every five minutes. It started at 19:35 at 100% battery and died at 20:55, so it lasted 1 hour 20 minutes.
- Failed AI battery test: A real AI workload for the battery test failed outright, so this is a CPU-and-memory stress figure. It says nothing about normal battery life.
- Noise: He says the fan was loud enough to hear from the next room under that load.
- Lenovo's claim: Lenovo gives no overall battery life figure. It only says 15 minutes of Rapid Charge Express gives three hours of capacity.
Hardware notes
Gewirtz describes a 2880 x 1800 OLED touchscreen and Wi-Fi 7 with Bluetooth 5.4. Ports are two USB-A (up to 10Gbps), two USB-C, full-size HDMI, a headphone jack and an SD card reader. He reports 17.5mm thickness and 4.28 pounds, which he says is nearly half a pound lighter than a 16-inch MacBook Pro. StorageReview also lists a 99.9Wh battery in a 4.28-pound chassis.
His one real hardware complaint is the pen. It lives in a removable magnetic holder that attaches only to the back of the lid. The holder can be knocked off, the pen must be taken out to charge, and he says the pen costs about $100 separately.
What this means for Windows developers
- Local AI coding works, with limits. Memory decides what you can run and how much context you get. Windows overhead takes a real share of 64GB.
- Check the configuration before buying. Fast prompt processing doesn't guarantee fast code generation. StorageReview found 32GB of GPU memory ran smaller models at full speed. It also found a roughly 40GB-class model won't fit on the Yoga.
- Reviews of this model differ. The Gadgeteer found the AI speed lead changed with the model: the Lenovo was faster on some, and an ASUS finished the Qwen and image-generation jobs sooner. One reviewer's results are not a universal ranking.
- Cloud tools still have an edge on context. A 32K window forces constant summarizing and reloading. For large codebases, that gap is as important as raw speed.
The most useful result is the combination: a thin Windows laptop produced a recognizable WPF-then-WinUI 3 prototype from foreign source with no cloud. But a 64GB cap, frequent context resets and a loud fan under stress show that running locally doesn't mean running smoothly. Gewirtz's own suggestion is a 128GB Yoga 9n option, which Lenovo doesn't currently offer in the 2-in-1.