NVIDIA RTX Spark

Other

NVIDIA RTX Spark: N1X superchip system (Blackwell RTX GPU + 20-core CPU) with LPDDR5X unified memory, SKUs from 24 to 128GB. Laptops start at $2,399 (ASUS / Dell / HP / Lenovo / Microsoft); the 140W desktop dev kit is $5,999.99. VRAM is logged at advertised capacity. Under Windows the memory is currently a fixed split, with roughly 62.5GB GPU-usable on the 128GB unit; **Linux has unlocked more**: a 2026-10-19 community hands-on (ASUS ProArt P16 128GB, Arch Linux ARM + NVIDIA 615.78 driver) reports the same 116.0 GiB unified memory pool to both the OS and CUDA, with CPU performance on par with Windows (7-Zip within ±1.2%). StorageReview measured STREAM Triad at 120GB/s (used for the bandwidth field), Llama-3.1-8B Q4_K_M at 56.5 tok/s and Qwen3.6-35B-A3B at 89.5 tok/s. Laptop TDP 45-80W, desktop 140W.

132

Run records

Models covered

Frameworks covered

L2 Cross-framework verified × 132

Evidence mix

Typical measured performance

Published benchmark averages per model × framework × quantization on this hardware.

llama.cpp · IQ3_S Decode 105.34 · Prefill 354.91 ·
Ollama · IQ2_M Decode 116.68 · Prefill 432.72 ·
llama.cpp · IQ3_XS Decode 72.77 · Prefill 239.26 ·
Ollama · IQ3_XS Decode 66.41 · Prefill 229.61 ·
llama.cpp · Q4_K_M Decode 82.35 · Prefill 342.05 ·
Ollama · IQ3_S Decode 107.47 · Prefill 568.49 ·
llama.cpp · Q4_K_M Decode 80.69 · Prefill 270.65 ·
Ollama · Q4_K_M Decode 79.19 · Prefill 309.44 ·
llama.cpp · IQ3_S Decode 91.56 · Prefill 457.14 ·
llama.cpp · Q3_K_S Decode 100.54 · Prefill 475.16 ·
llama.cpp · Q4_K_M Decode 76.22 · Prefill 300.34 ·
Ollama · IQ4_XS Decode 74.46 · Prefill 328.19 ·
Ollama · Q2_K Decode 96.77 · Prefill 510.91 ·
Ollama · Q4_K_M Decode 70.03 · Prefill 271.9 ·
llama.cpp · Q4_K_M Decode 47.78 · Prefill 218.5 ·
Ollama · IQ4_XS Decode 49.65 · Prefill 223.79 ·
Ollama · Q4_K_M Decode 45.05 · Prefill 148.34 ·
llama.cpp · IQ1_S Decode 15.77 · Prefill 62.48 ·
llama.cpp · Q4_K_M Decode 11.03 · Prefill 58.41 ·
llama.cpp · Q4_K_S Decode 11.75 · Prefill 39.19 ·
Ollama · IQ3_S Decode 12.19 · Prefill 51.57 ·
Ollama · IQ4_XS Decode 9.95 · Prefill 46.04 ·
Ollama · Q4_K_M Decode 9.21 · Prefill 45.35 ·
llama.cpp · IQ4_XS Decode 53.68 · Prefill 258.94 ·
llama.cpp · Q4_K_M Decode 50.55 · Prefill 236.22 ·
Ollama · Q4_K_M Decode 49.53 · Prefill 221.38 ·
llama.cpp · Q5_K_M Decode 8.81 · Prefill 37.51 ·
llama.cpp · Q8_0 Decode 6.42 · Prefill 24.36 ·
Ollama · Q5_K_M Decode 7.6 · Prefill 28.29 ·
Ollama · Q6_K Decode 7.47 · Prefill 38.07 ·
Ollama · Q8_0 Decode 6.14 · Prefill 31.39 ·
llama.cpp · FP16 Decode 5.5 · Prefill 19.76 ·
llama.cpp · Q5_K_M Decode 8.41 · Prefill 32.22 ·
llama.cpp · Q8_0 Decode 7.68 · Prefill 32.82 ·
Ollama · Q5_K_M Decode 8.83 · Prefill 30.38 ·
Ollama · Q8_0 Decode 6.72 · Prefill 25.96 ·
llama.cpp · Q5_K_M Decode 78.77 · Prefill 418.38 ·
llama.cpp · Q8_0 Decode 57.81 · Prefill 190.18 ·
Ollama · Q8_0 Decode 52.36 · Prefill 271.04 ·
llama.cpp · Q4_K_M Decode 90.26 · Prefill 398.49 ·
llama.cpp · Q5_K_M Decode 71.83 · Prefill 346.83 ·
llama.cpp · Q8_0 Decode 64.66 · Prefill 241.03 ·
Ollama · FP16 Decode 40.67 · Prefill 196.29 ·
Ollama · Q5_K_M Decode 69.06 · Prefill 289.88 ·
Ollama · Q8_0 Decode 56.21 · Prefill 286.64 ·
llama.cpp · Q6_K Decode 8.03 · Prefill 30.64 ·
llama.cpp · Q8_0 Decode 6.76 · Prefill 25.64 ·
Ollama · FP16 Decode 4.9 · Prefill 24.42 ·
Ollama · Q6_K Decode 7.03 · Prefill 31.25 ·
Ollama · Q8_0 Decode 6.87 · Prefill 25.86 ·

Fit check

Physics-based estimate from VRAM and memory bandwidth: pick a model to see which quantizations fit and the expected speed

Formula: tok/s = (bandwidth ÷ model size) × 0.55, adapted from llmfit; see run records for measured numbers

Key specs

Vendor: NVIDIA
Architecture: N1X (Blackwell RTX GPU + 20-core CPU)
VRAM: 128 GB
Memory bandwidth: 120 GB/s
TDP: 80 W
Release date: 2026-10-16

Run records

50 configurations

Model	Framework	Quantization	Decode	Gen time	Samples
gemma-4-E4B-it	llama.cpp	IQ3_S	105.3 tok/s	—	1
gemma-4-E4B-it	Ollama	IQ2_M	116.7 tok/s	—	1
Ministral-3-8B-Instruct-2512	llama.cpp	IQ3_XS	72.8 tok/s	—	1
Ministral-3-8B-Instruct-2512	Ollama	IQ3_XS	66.4 tok/s	—	1
iFlytek Spark-X2.5-4B	llama.cpp	Q4_K_M	82.3 tok/s	—	1
iFlytek Spark-X2.5-4B	Ollama	IQ3_S	107.5 tok/s	—	1
NVIDIA-Nemotron-3.5-Bolt-E4B-4GB	llama.cpp	Q4_K_M	80.7 tok/s	—	1
NVIDIA-Nemotron-3.5-Bolt-E4B-4GB	Ollama	Q4_K_M	79.2 tok/s	—	1
Flint-40B-A2.5B	llama.cpp	IQ3_S	91.6 tok/s	—	1
Flint-40B-A2.5B	llama.cpp	Q3_K_S	100.5 tok/s	—	1
Flint-40B-A2.5B	llama.cpp	Q4_K_M	76.2 tok/s	—	1
Flint-40B-A2.5B	Ollama	IQ4_XS	74.5 tok/s	—	1
Flint-40B-A2.5B	Ollama	Q2_K	96.8 tok/s	—	1
Flint-40B-A2.5B	Ollama	Q4_K_M	70 tok/s	—	1
Raon-1-mini-preview	llama.cpp	Q4_K_M	47.8 tok/s	—	2
Raon-1-mini-preview	Ollama	IQ4_XS	49.6 tok/s	—	1
Raon-1-mini-preview	Ollama	Q4_K_M	45 tok/s	—	1
mammoth-1	llama.cpp	IQ1_S	15.8 tok/s	—	1
mammoth-1	llama.cpp	Q4_K_M	11 tok/s	—	1
mammoth-1	llama.cpp	Q4_K_S	11.8 tok/s	—	1
mammoth-1	Ollama	IQ3_S	12.2 tok/s	—	1
mammoth-1	Ollama	IQ4_XS	9.9 tok/s	—	1
mammoth-1	Ollama	Q4_K_M	9.2 tok/s	—	1
elephant-1	llama.cpp	IQ4_XS	53.7 tok/s	—	1
elephant-1	llama.cpp	Q4_K_M	50.5 tok/s	—	2
elephant-1	Ollama	Q4_K_M	49.5 tok/s	—	2
Qwen3.8-27B	llama.cpp	Q5_K_M	8.8 tok/s	—	2
Qwen3.8-27B	llama.cpp	Q8_0	6.4 tok/s	—	1
Qwen3.8-27B	Ollama	Q5_K_M	7.6 tok/s	—	1
Qwen3.8-27B	Ollama	Q6_K	7.5 tok/s	—	1
Qwen3.8-27B	Ollama	Q8_0	6.1 tok/s	—	1
Qwen3.6-27B	llama.cpp	FP16	5.5 tok/s	—	1
Qwen3.6-27B	llama.cpp	Q5_K_M	8.4 tok/s	—	1
Qwen3.6-27B	llama.cpp	Q8_0	7.7 tok/s	—	1
Qwen3.6-27B	Ollama	Q5_K_M	8.8 tok/s	—	1
Qwen3.6-27B	Ollama	Q8_0	6.7 tok/s	—	1
Qwen3.6-35B-A3B	llama.cpp	Q5_K_M	78.8 tok/s	—	1
Qwen3.6-35B-A3B	llama.cpp	Q8_0	57.8 tok/s	—	1
Qwen3.6-35B-A3B	Ollama	Q8_0	52.4 tok/s	—	1
Qwen3.5-4B	llama.cpp	Q4_K_M	90.3 tok/s	—	2
Qwen3.5-4B	llama.cpp	Q5_K_M	71.8 tok/s	—	1
Qwen3.5-4B	llama.cpp	Q8_0	64.7 tok/s	—	1
Qwen3.5-4B	Ollama	FP16	40.7 tok/s	—	1
Qwen3.5-4B	Ollama	Q5_K_M	69.1 tok/s	—	2
Qwen3.5-4B	Ollama	Q8_0	56.2 tok/s	—	1
Qwen3.5-27B	llama.cpp	Q6_K	8 tok/s	—	1
Qwen3.5-27B	llama.cpp	Q8_0	6.8 tok/s	—	1
Qwen3.5-27B	Ollama	FP16	4.9 tok/s	—	1
Qwen3.5-27B	Ollama	Q6_K	7 tok/s	—	1
Qwen3.5-27B	Ollama	Q8_0	6.9 tok/s	—	1