The Lenovo Yoga 9n 2-in-1 ran local coding workloads in two reported tests, but its most ambitious app-porting demonstration produced a rough Windows facsimile, not finished software. In David Gewirtz’s ZDNET test, OpenCode and Qwen reached that result in about four hours; StorageReview’s separate fixed coding task completed successfully in two of three runs.
What the local coding demo produced
Gewirtz described his test laptop as having a 20-core N1X processor, 64GB of memory and a 1TB SSD. He ran OpenCode v2.1.4 with Qwen3-Coder-Next 80B-A3B, quantized to Q4_K_M, using llama.cpp with NVIDIA CUDA.
The task was to turn a Mac-only icon-gallery app into a Windows version. OpenCode moved the project from Swift to C#, first using WPF and then WinUI 3. After about four hours, the app rendered icons and roughly imitated the original. More work remained before the Windows version could be distributed.
The setup’s largest working context window was 32,768 tokens. A context window is the amount of text a model can consider at once. Gewirtz said the agent typically handled one or two actions before it had to save its status and refresh the context, interrupting the flow of work.
A separate coding benchmark hit a memory limit
StorageReview tested a different Yoga 9n 2-in-1, model 83N7000UUS, with a 20-core CPU, 64GB of memory and a 1TB SSD. In its fixed Codex CLI tier-S task using gpt-oss:20b, the laptop averaged 58.0 tokens per second and a median completion time of 132.5 seconds. It succeeded in two of three runs.
That test allocated 32GB of the system’s 64GB of unified memory to the GPU and 32GB to system use. StorageReview did not run its tier-M Qwen3.6-35B-A3B task because that workload called for about 40GB of GPU memory. The result is a practical reminder for local-AI work: total system memory and the amount allocated to the GPU are not interchangeable.
This fixed benchmark supports a different conclusion from the app-porting demo. The Yoga handled a defined coding task at a measured rate, but the task did not succeed on every run, and the larger workload exceeded the test system’s GPU-memory allocation.
Prompt processing and token generation tell different stories
In separate llama.cpp tests, Gewirtz compared the Yoga with a 2025 Mac Studio equipped with an M4 Max, 128GB of unified memory and an 8TB SSD. The Yoga processed prompts faster: 2,647 versus 899 input tokens per second for the 8B model, and 628 versus 203 for the 32B model. The Mac generated output faster: 82 versus 53 output tokens per second for 8B, and 23 versus 9 for 32B.
Prompt processing and output generation are different stages of model inference, so these figures do not produce one overall speed ranking. Gewirtz also described the comparison as not entirely fair: the Mac had twice the Yoga’s memory and eight times its listed storage capacity.
Sustained performance and temperature
In StorageReview’s roughly 30-minute Cinebench 2026 loop, the Yoga recorded a steady-state score of 6,256, or 100.1% of its first pass, with a reported maximum temperature of 72°C. In the same review’s comparison, an HP OmniBook Ultra 16 scored 6,942 and reached 98°C. The Yoga ran cooler in that test, while the HP posted the higher sustained score; the comparison involved different laptop configurations.
Lenovo’s listed configurations
Lenovo’s Yoga 9n 2-in-1 16N1X11 family pairs its N1X processor options with soldered unified memory:
| Configuration | CPU cores | Blackwell GPU cores | Unified memory | Upgradeable memory |
| 18-core | 18 | 5,120 | 32GB | No, soldered |
| 20-core | 20 | 6,144 | 64GB | No, soldered |
Lenovo also lists a 16-inch 2,880 × 1,800 OLED touchscreen with a refresh rate of up to 120Hz and a 99.9Wh battery. The model was announced on September 3, 2026, as part of Lenovo’s RTX Spark lineup. NeoTeo’s earlier overview of the announced Yoga models covers that launch context.
Because the memory is soldered, the capacity choice is made at configuration time. For local AI, check the GPU memory allocation required by the workload, not just the laptop’s total memory: StorageReview’s 64GB test system assigned 32GB to GPU use, below the roughly 40GB needed for its larger task.
What the 80-minute battery run measured
The 80-minute result in Gewirtz’s review came from a custom CPU-and-memory stress script, not an AI coding session. His attempted AI workload failed; the replacement script ran 18 CPU processes doing SHA-256 and OpenSSL work, plus a memory-churning process. The laptop went from 100% charge at 19:35 to shutdown at 20:55.
Lenovo lists a 99.9Wh battery and an approximate maximum of 19.83 hours for local 1080p video playback at 150 nits. That manufacturer figure describes a different workload and test condition from Gewirtz’s stress run.
For developers whose local coding tasks fit the tested GPU-memory allocation, the Yoga has demonstrated useful capability, though the fixed benchmark succeeded in only two of three attempts. Workloads needing around 40GB of GPU memory exceeded the allocation in StorageReview’s 64GB configuration, and Lenovo does not allow its memory to be upgraded.
