Builds / Nvidia RTX Spark N1X, 6,144-core Blackwell GPU (unified)
Lenovo Yoga Pro 9n (RTX Spark, 128 GB): gpt-oss 20B at 80.4 tok/s, Qwen3.6 35B-A3B at 72.1 (llama.cpp CUDA, Windows on Arm)
Last updated:
- GPU(s)
- Nvidia RTX Spark N1X, 6,144-core Blackwell GPU (unified)
- VRAM / unified memory (GB)
- 128
- CPU
- Nvidia N1X (RTX Spark), 20 Arm cores
Amazon · eBay (used) · Newegg - Motherboard
- Lenovo Yoga Pro 9n laptop
Amazon · eBay (used) · Newegg - RAM
- 128 GB LPDDR5X-9400 unified (64 GB reserved for graphics by the NVIDIA app by default)
- PSU
- not stated
- Other
- Windows 11 Home on Arm, timed on mains power; llama.cpp b11669 CUDA build for Windows on Arm, whole model on the GPU
- Reported total cost
- not stated
- Power
- not stated
Measured speed (generation tokens/sec)
| Model | Quant | Engine | tok/s | Prompt processing (tok/s) | Details |
|---|---|---|---|---|---|
| gpt-oss 20B | MXFP4 (12.1 GB GGUF), reasoning low | llama.cpp b11669 (CUDA, Windows on Arm) | 80.4 | 4,415 @ 15,000-token prompt est. TTFT @8k ≈ 1.9 s (calculated from measured prefill) | context: code prompt, 32K window📏 measuredlocally test: one user, thinking off, same code prompt on every machine; middle of three runs (idle machine, 8 quiet seconds before each run); llama.cpp b11669, 32,768-token window; 78.2 in WSL 2 |
| Qwen3.6 35B-A3B | UD-Q4_K_M (22.1 GB) | llama.cpp b11669 (CUDA, Windows on Arm) | 72.1 | 2,339 @ 19,000-token prompt est. TTFT @8k ≈ 3.5 s (calculated from measured prefill) | context: code prompt, 32K window📏 measuredlocally test: one user, thinking off, same code prompt on every machine; middle of three runs (idle machine, 8 quiet seconds before each run); llama.cpp b11669, 32,768-token window; 69.7 in WSL 2 |
| Gemma 4 26B-A4B | UD-IQ4_XS (13.6 GB) | llama.cpp b11669 (CUDA, Windows on Arm) | 60.7 | 2,328 @ 19,000-token prompt est. TTFT @8k ≈ 3.5 s (calculated from measured prefill) | context: code prompt, 32K window📏 measuredlocally test: one user, thinking off, same code prompt on every machine; middle of three runs (idle machine, 8 quiet seconds before each run); llama.cpp b11669, 32,768-token window; 58.3 in WSL 2 |
| Qwen3.5 9B | Q4_K_M (5.7 GB) | llama.cpp b11669 (CUDA, Windows on Arm) | 44.6 | 2,684 @ 19,000-token prompt est. TTFT @8k ≈ 3.1 s (calculated from measured prefill) | context: code prompt, 32K window📏 measuredlocally test: one user, thinking off, same code prompt on every machine; middle of three runs (idle machine, 8 quiet seconds before each run); llama.cpp b11669, 32,768-token window; 40.5 in WSL 2 |
| Gemma 4 12B | QAT Q4_0 (7.0 GB, Google's file) | llama.cpp b11669 (CUDA, Windows on Arm) | 33.2 | 1,513 @ 19,000-token prompt est. TTFT @8k ≈ 5.4 s (calculated from measured prefill) | context: code prompt, 32K window📏 measuredlocally test: one user, thinking off, same code prompt on every machine; middle of three runs (idle machine, 8 quiet seconds before each run); llama.cpp b11669, 32,768-token window; 28.6 in WSL 2 |
| Qwen3.8 27B | Q4_K_M (17.1 GB, Unsloth) | llama.cpp b11669 (CUDA, Windows on Arm) | 15.2 | 800 @ 19,000-token prompt est. TTFT @8k ≈ 10.2 s (calculated from measured prefill) | context: code prompt, 32K window📏 measuredlocally test: one user, thinking off, same code prompt on every machine; middle of three runs (idle machine, 8 quiet seconds before each run); llama.cpp b11669, 32,768-token window; 14.1 in WSL 2 |
| Qwen3.8 27B | Q4_K_M (17.1 GB, Unsloth) + MTP (2 draft tokens) | llama.cpp b11669 (CUDA, Windows on Arm) | 33.3 | not stated | ⚡ spec-decode (MTP)context: code prompt, 32K window📏 measuredlocally test: one user, thinking off, same code prompt on every machine; middle of three runs (idle machine, 8 quiet seconds before each run); llama.cpp b11669, 32,768-token window; 24.5 on the prose prompt, 30.7 code in WSL 2 |
Prompt processing = prefill speed (how fast the prompt is read) as reported by the source, with the prompt length it was measured at; blank = the source does not report it. est. TTFT @8k is calculated from the measured prefill speed (8,192 ÷ prefill tok/s), shown only when prefill was measured on a prompt of 512+ tokens. It is not a measurement and ignores load time and caching. measured TTFT is the source's own time-to-first-token. Why prefill matters for coding assistants.
measuredlocally.com (Alex, UK), tested 29 Oct 2026 on his own laptop. Native Windows rows listed; the same tests inside WSL 2 were slower on every model (noted per row). With 16 users at once Qwen3.5 9B reached 188 tok/s in total. A ~2-bit Nemotron 3.5 Bolt run (41.5 tok/s) is below Q4 and not listed.
https://measuredlocally.com/hardware-leaderboard.html
Blank fields = not stated in the source. We never estimate build data; GPU street prices above are separately sourced and dated.
Copy into a README or blog. Links back to this build. See badge docs (CC BY 4.0 attribution). The SVG does not invent speeds — numbers stay on the build page.
Markdown
[](https://getrigbase.com/builds/lenovo-yoga-pro-9n-rtx-spark-128gb-measuredlocally/)
HTML
<a href="https://getrigbase.com/builds/lenovo-yoga-pro-9n-rtx-spark-128gb-measuredlocally/" rel="noopener"><img src="https://getrigbase.com/badge/lenovo-yoga-pro-9n-rtx-spark-128gb-measuredlocally.svg" alt="Measured on RigBase" height="20"/></a>
Preview: 
· How to use
Quick answers: Can RTX Spark N1X run Gemma 4 12B? · Can RTX Spark N1X run Gemma 4 26B-A4B? · Can RTX Spark N1X run Qwen3.5 9B? · Can RTX Spark N1X run Qwen3.6 35B-A3B? · Can RTX Spark N1X run Qwen3.8 27B? · Can RTX Spark N1X run gpt-oss 20B?