Builds / Nvidia RTX Spark N1X, 6,144-core Blackwell GPU (unified)
Surface Laptop Ultra (RTX Spark N1X, 128 GB): 8 models on llama.cpp b12110, Flint 104.8 tok/s, gpt-oss 120B 64.2, Qwen3.8 27B 16.5 (maruhana)
Last updated:
- GPU(s)
- Nvidia RTX Spark N1X, 6,144-core Blackwell GPU (unified)
- VRAM / unified memory (GB)
- 82
- CPU
- Nvidia RTX Spark N1X, 20-core Grace CPU
Amazon Β· eBay (used) Β· Newegg - Motherboard
- Microsoft Surface Laptop Ultra
Amazon Β· eBay (used) Β· Newegg - RAM
- 128 GB unified, 83,968 MiB assigned to the GPU
- PSU
- not stated
- Other
- Windows 11 Arm64; llama.cpp b12110 CUDA build; other GPU workloads stopped during measurement
- Reported total cost
- not stated
- Power
- not stated
Measured speed (generation tokens/sec)
| Model | Quant | Engine | tok/s | Prompt processing (tok/s) | Details |
|---|---|---|---|---|---|
| Flint 40B-A2.5B | NVFP4 (22.9 GB) | llama.cpp b12110 (CUDA, Windows on Arm) | 104.8 | 2,066 @ 512-token prompt est. TTFT @8k β 4.0 s (calculated from measured prefill) | context: llama-benchπ llama-bench tg128, mean of 3 runs (std dev 0.7 or less), -ngl 999 -fa 1, speculative decoding off; reading a 32,768-token prompt 1,582 tok/s (1 run) |
| Qwen3.6 35B-A3B | UD-Q4_K_M (22.1 GB) | llama.cpp b12110 (CUDA, Windows on Arm) | 75.7 | 3,092 @ 512-token prompt est. TTFT @8k β 2.6 s (calculated from measured prefill) | context: llama-benchπ llama-bench tg128, mean of 3 runs (std dev 0.7 or less), -ngl 999 -fa 1, speculative decoding off; reading a 32,768-token prompt 2,138 tok/s (1 run); 4 requests at once 110.0 total |
| gpt-oss 120B | MXFP4 (63.4 GB) | llama.cpp b12110 (CUDA, Windows on Arm) | 64.2 | 2,116 @ 512-token prompt est. TTFT @8k β 3.9 s (calculated from measured prefill) | context: llama-benchπ llama-bench tg128, mean of 3 runs (std dev 0.7 or less), -ngl 999 -fa 1, speculative decoding off; reading a 32,768-token prompt 1,708 tok/s (1 run) |
| Nemotron 3.5 Bolt 74B-A9B | NVIDIA 24GB GGUF, experts IQ2_XXS (22.0 GB) | llama.cpp b12110 (CUDA, Windows on Arm) | 42.7 | 1,174 @ 512-token prompt est. TTFT @8k β 7.0 s (calculated from measured prefill) | context: llama-benchπ llama-bench tg128, mean of 3 runs (std dev 0.7 or less), -ngl 999 -fa 1, speculative decoding off; reading a 32,768-token prompt 1,083 tok/s (1 run); with the official LiLiCorr drafter (--spec-type draft-dflash, 7 drafts) short answers ran at 105-115 tok/s, longer Japanese answers 55; 4 requests at once 87.1 total |
| Nemotron 3.5 Super VL 120B-A12B | UD-Q3_K_XL (65.8 GB) | llama.cpp b12110 (CUDA, Windows on Arm) | 26.6 | 635 @ 512-token prompt est. TTFT @8k β 12.9 s (calculated from measured prefill) | context: llama-benchπ llama-bench tg128, mean of 3 runs (std dev 0.7 or less), -ngl 999 -fa 1, speculative decoding off; reading a 32,768-token prompt 596 tok/s (1 run); UD-Q4_K_XL (86.9 GB) also ran fully offloaded at 23.9 (pp512 625.1, 1 run) |
| Qwen3.8 27B | UD-Q4_K_M (16.5 GB) | llama.cpp b12110 (CUDA, Windows on Arm) | 16.5 | 942 @ 512-token prompt est. TTFT @8k β 8.7 s (calculated from measured prefill) | context: llama-benchπ llama-bench tg128, mean of 3 runs (std dev 0.7 or less), -ngl 999 -fa 1, speculative decoding off; reading a 32,768-token prompt 836 tok/s (1 run) |
| Mistral Small 3.2 24B legacy | Q4_K_M (14.3 GB) | llama.cpp b12110 (CUDA, Windows on Arm) | 18.5 | 1,183 @ 512-token prompt est. TTFT @8k β 6.9 s (calculated from measured prefill) | context: llama-benchπ llama-bench tg128, mean of 3 runs (std dev 0.7 or less), -ngl 999 -fa 1, speculative decoding off; reading a 32,768-token prompt 952 tok/s (1 run) |
| Hermes 4.3 36B legacy | Q4_K_M (21.8 GB) | llama.cpp b12110 (CUDA, Windows on Arm) | 12.0 | 733 @ 512-token prompt est. TTFT @8k β 11.2 s (calculated from measured prefill) | context: llama-benchπ llama-bench tg128, mean of 3 runs (std dev 0.7 or less), -ngl 999 -fa 1, speculative decoding off; reading a 32,768-token prompt 393 tok/s (1 run) |
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 Γ· 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.
Owner-run test by maruhana (dev.to, 14 Nov 2026; English adaptation of their Zenn article), same build and commands for all 8 models; quantization was not aligned across models. A config-fixing agent task was solved by 6 of 8 models (Flint and Hermes 4.3 did not fix it in 12 steps). Compare with Micro Center's Surface Laptop Ultra numbers (Qwen3.8 27B 14.5).
Blank fields = not stated in the source. We never estimate build data; GPU street prices above are separately sourced and dated.
Quick answers: Can RTX Spark N1X run Flint 40B-A2.5B? Β· Can RTX Spark N1X run Nemotron 3.5 Bolt 74B-A9B? Β· Can RTX Spark N1X run Nemotron 3.5 Super VL 120B-A12B? Β· Can RTX Spark N1X run Qwen3.6 35B-A3B? Β· Can RTX Spark N1X run Qwen3.8 27B? Β· Can RTX Spark N1X run gpt-oss 120B?