RTX Spark N1X (Windows ARM64) — Qwen3.8-Flash-Next 176.94B IQ4_XS llama.cpp Benchmark

[knmun]

RTX Spark N1X (Windows ARM64) — Qwen3.8-Flash-Next 176.94B IQ4_XS llama.cpp Benchmark

I benchmarked Qwen3.8-Flash-Next on a single NVIDIA RTX Spark N1X running Windows ARM64 using llama.cpp with the CUDA backend.

## Hardware and software

• GPU: NVIDIA RTX Spark N1X (6144-core Blackwell GPU, compute capability 12.1)

• Memory: 128GB unified memory

• OS: Windows ARM64

• llama.cpp: build 12024, commit ddf48f85c

• Model: Qwen3.8-Flash-Next, 176.94B total parameters, A3B

• Quantization: UD-IQ4_XS (4.25 bpw)

• Model size: 87.24 GiB

• Flash Attention: Enabled

• KV Cache: Q8_0 / Q8_0

• GPU offloading: 999 layers

• Batch / Micro-batch: 2048 / 512

• Split mode: none

• Repetitions: 5

Benchmark results

Test	Average tok/s	Std. deviation
Prefill 2048	313.36	±162.68
Prefill 8192	587.59	±4.58
Decode 128	30.83	±0.32
Decode 512	30.97	±0.05

The 2048-token prefill results varied considerably across runs (167.61–592.86 tok/s), while the 8192-token prefill and decode results were stable.

The sustained decode throughput was approximately 31 tokens/s, without speculative decoding or MTP.

Benchmark command

.\llama-bench.exe `
  -m "Qwen3.8-Flash-Next-UD-IQ4_XS-00001-of-00003.gguf" `
  -ngl 999 `
  -p 2048,8192 `
  -n 128,512 `
  -b 2048 `
  -ub 512 `
  -fa 1 `
  -ctk q8_0 `
  -ctv q8_0 `
  -sm none `
  -r 5 `
  -o json

The first file of the three-part GGUF model was used to load the model.

These are llama-bench results at context depth 0, not end-to-end server throughput measurements. Longer-context benchmarks have not yet been included.

I would be interested in comparing these results with DGX Spark Linux and other Windows ARM64 configurations.

spark-bench.json